Author: Paula Ramos, PhD

Welcome to the final installment of our Cosmos Edition of “The Best of CVPR.” Over the past four parts, we explored physical reasoning benchmark, embodied robotics, driving simulation, and video generation. Now we close the series by looking at the infrastructure layer, the tokenizers, datasets, representations, and training strategies that make everything else possible.

These are the papers that don’t always make headlines, but they’re the ones that enable the breakthroughs. Better tokenizers mean more efficient world models. Better datasets mean more grounded training. Better representations mean more generalizable systems. Together, they form the foundation of the Physical AI stack.

Rethinking How We Tokenize Video — AdapTok [1]

Paper Title: AdapTok: Learning Adaptive and Temporally Causal Video Tokenization in a 1D Latent Space
Paper Authors: Y. Li, C. Tian, R. Xia, N. Liao, W. Guo, J. Yan, H. Li, et al.
Institution: Shanghai Jiao Tong University, The Chinese University of Hong Kong, Tsinghua University, Tongji University, NVIDIA, Shanghai AI Laboratory

What if video tokenizers could adapt, using more tokens for complex scenes and fewer for simple ones, while maintaining causal processing for real-time applications?

What It’s About

AdapTok introduces an adaptive, temporally causal video tokenizer that represents videos with variable-length latent tokens in a 1D space. Unlike fixed-resolution tokenizers that use the same number of tokens for every frame regardless of complexity, AdapTok allocates tokens based on scene content, enabling more efficient and flexible representations.

Why It Matters

Video tokenization is the gateway to everything in video world modeling, the quality of the tokenizer directly determines what downstream models can learn. NVIDIA’s Cosmos Tokenizer pioneered causal temporal tokenization using causal convolution and attention layers. AdapTok builds on this foundation by adding adaptivity: scenes with more motion or detail get more tokens, while static scenes get fewer. This efficiency gain cascades through the entire pipeline.

How It Works

• Adopts a transformer-based architecture with causal attention masks for temporal modeling, ensuring each frame only depends on past frames (not future ones).
• Introduces an adaptive token allocation mechanism that adjusts the number of latent tokens per frame based on scene complexity.
• Builds explicitly on the Cosmos Tokenizer’s approach to causal convolution, extending it with adaptive capabilities.
• Benchmarks against Cosmos-Tokenizer-DV and other baselines (OmniTokenizer, ElasticTok) on reconstruction quality (rFVD).
• Achieves competitive quality with significantly fewer tokens on average.

Key Result

AdapTok achieves lower rFVD than Cosmos-Tokenizer-DV while using fewer tokens on average, particularly excelling on videos with varying complexity. The adaptive allocation means the tokenizer naturally spends capacity where it’s needed most, avoiding waste on static or low-detail segments.

Broader Impact

AdapTok validates the Cosmos Tokenizer as foundational infrastructure while pushing it further. The adaptive approach could significantly reduce the computational cost of training and running video world models, making them more practical for real-time robotics, streaming applications, and edge deployment. This work from SJTU, CUHK, Tsinghua, and NVIDIA demonstrates the kind of incremental-but-critical improvement that moves the field forward.

One Million Videos, Fully Annotated — SceneScribe-1M [2]

Paper Title: SceneScribe-1M: A Large-Scale Video Dataset with Comprehensive Geometric and Semantic Annotations
Paper Authors: Y. Wang, K. Zheng, J. Wang, M. Chen, D. Novotny, et al.
Institution: Shanghai Jiao Tong University, Ant Group, University of Oxford

AI models are only as good as their training data. What if we had a million in-the-wild videos, each annotated with text descriptions, camera parameters, dense depth maps, and consistent 3D point tracks?

What It’s About

SceneScribe-1M is a new large-scale, multi-modal video dataset comprising one million in-the-wild videos, each meticulously annotated with detailed textual descriptions, precise camera parameters, dense depth maps, and consistent 3D point tracks. It bridges the gap between 3D geometric perception and video synthesis, providing a unified resource for both domains.

Why It Matters

The convergence of 3D understanding and video generation has created demand for datasets that support both, but existing resources address one or the other, not both simultaneously. SceneScribe-1M fills this gap, enabling research on world foundation models that need to both perceive the 3D structure of scenes and generate realistic video. It cites the Cosmos World Foundation Model Platform as a key motivator for this kind of unified data.

How It Works

• Collects one million in-the-wild videos from diverse sources.
• Generates comprehensive annotations using automated pipelines: textual descriptions, camera intrinsics and extrinsics, per-pixel dense depth maps, and temporally consistent 3D point tracks.
• Establishes benchmarks across multiple downstream tasks: monocular depth estimation, scene reconstruction, dynamic point tracking, and text-to-video synthesis with and without camera control.
• Designed to support world foundation models that need both geometric understanding and generative capabilities.

Key Result

Models trained on SceneScribe-1M demonstrate strong performance across all benchmark tasks, validating that the automated annotations are high-quality enough to train competitive systems. The dataset shows particular value for tasks at the intersection of 3D understanding and video generation, exactly where world models operate.

Broader Impact

SceneScribe-1M from SJTU, Ant Group, and Oxford addresses a fundamental bottleneck: world models need training data with both geometric and semantic annotations at scale. By open-sourcing this dataset, the authors provide a catalyst for the entire field of world foundation modeling. The explicit connection to Cosmos signals that this work is designed to enable the next generation of physical AI systems.

3D Registration Without Training — C-GenReg [3]

Paper Title: C-GenReg: Training-Free 3D Point Cloud Registration by Multi-View-Consistent Geometry-to-Image Generation with Probabilistic Modalities Fusion
Paper Authors: Y. Haitman, A. Efraim, J. M. Francos
Institution: Ben-Gurion University of the Negev

What if you could align two 3D point clouds without any training at all, by turning them into images using a world model and then matching those images?

What It’s About

C-GenReg introduces a training-free framework for 3D point cloud registration that leverages Cosmos-Transfer as its core component. The idea is creative: render depth maps from the input point clouds, use Cosmos-Transfer to generate multi-view-consistent RGB images from those depth maps, then use standard image matching techniques to find correspondences and compute the alignment.

Why It Matters

Point cloud registration, aligning two 3D scans of the same scene, is a fundamental problem in 3D vision, robotics, and autonomous driving. Most methods require training on specific datasets, which limits generalization. C-GenReg avoids this entirely by relying on the world priors learned by Cosmos-Transfer to generate consistent images from depth inputs, then solving registration in the image domain where mature matching algorithms exist.

How It Works

• Renders depth maps from each input point cloud using virtual cameras.
• Uses Cosmos-Transfer WFM to generate multi-view-consistent RGB images from the depth sequences, leveraging its strong world priors for geometric coherence and appearance consistency.
• Computes correspondences using both geometric (from the original point clouds) and image-derived (from the generated RGB images) features.
• Fuses these modality-specific correspondences using a probabilistic match-then-fuse formulation.
• Requires zero training, the entire pipeline relies on pretrained models and geometric computation.

Key Result

C-GenReg achieves competitive registration accuracy on the 3DMatch benchmark with rotation errors as low as 0.68° and translation errors of 3.5cm — all without any task-specific training. The key insight: Cosmos-Transfer’s world priors produce images that are geometrically coherent and appearance-consistent, making them suitable for reliable feature matching.

Broader Impact

C-GenReg from Ben-Gurion University demonstrates a novel use case for world foundation models that goes far beyond video generation. By repurposing Cosmos-Transfer as a “geometry-to-image” translator, the paper shows that WFMs encode world knowledge that can be applied to geometric vision tasks. This opens the door for many other applications where turning structured data into rich visual representations could unlock new capabilities.

Learning 3D from Unlabeled Internet Videos — Lifting Unlabeled Data [4]

Paper Title: Lifting Unlabeled Internet-level Data for 3D Scene Understanding
Paper Authors: Y. Chen, Y. Zhang, H. Yu, J. He, Y. Wang, J. Huang, et al.
Institution: Peking University, Beijing University of Posts and Telecommunications, Tsinghua University, BIGAI

Annotated 3D data is scarce and expensive. But what if unlimited unlabeled internet videos could be turned into 3D training data automatically?

What It’s About

This paper demonstrates that carefully designed data engines can leverage web-curated, unlabeled videos to automatically generate training data for 3D scene understanding. The authors identify bottlenecks in automated data generation, reveal critical factors for effective learning from unlabeled data, and validate their approach across multiple perception granularities, from low-level 3D detection to high-level spatial VQA.

Why It Matters

The bottleneck in 3D scene understanding is data, not algorithms. Annotating 3D scenes with depth, instance labels, and spatial relationships is orders of magnitude more expensive than annotating images. This paper shows that by combining video understanding models with geometric reconstruction pipelines, you can generate 3D training data at internet scale, a game-changer for embodied AI and spatial intelligence.

How It Works

• Curates large-scale unlabeled videos from the internet.
• Applies automated pipelines to extract 3D annotations: depth maps, camera poses, object instances, and spatial relationships.
• Identifies and analyzes bottlenecks in the automated generation process, quality of depth estimation, accuracy of camera pose, consistency of instance segmentation.
• Evaluates on three tasks spanning different perception levels: 3D object detection, instance segmentation (low-level), and spatial VQA and vision-language navigation (high-level).
• Cites Cosmos-Predict2.5 and the Cosmos WFM Platform as part of the world foundation model landscape that motivates this data-centric approach.

Key Result

Models trained on the automatically generated data demonstrate strong zero-shot performance across all three tasks and show further improvement after fine-tuning on task-specific datasets. This validates that internet-scale unlabeled video can serve as a viable path toward more capable scene understanding systems.

Broader Impact

This work from Peking University, Tsinghua, and BIGAI validates a crucial hypothesis: the data problem in 3D understanding can be solved by leveraging the abundance of unlabeled internet video. As world foundation models like Cosmos push toward richer 3D representations, having scalable data engines that can produce training data without human annotation becomes essential infrastructure.

Teaching Robot Policies to Improve Themselves — SRPO [5]

Paper Title: SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models
Paper Authors: S. Fei, S. Wang, L. Ji, A. Li, S. Zhang, L. Liu, J. Hou, et al.
Institution: Fudan University, Tongji University

VLA models learn from demonstrations, but what happens when those demonstrations aren’t enough? What if the model could use a world model to imagine outcomes and optimize itself?

What It’s About

SRPO (Self-Referential Policy Optimization) is a novel reinforcement learning framework for VLA models that eliminates the need for external demonstrations or manual reward engineering. Instead of relying on binary success signals, SRPO uses the model’s own predictions to generate reference trajectories, optimizing the policy through self-referential feedback.

Why It Matters

Current VLA-RL methods suffer from severe reward sparsity, binary success/fail signals waste the information contained in failed trajectories. SRPO addresses this by creating richer training signals from the model’s own behavior. The paper specifically explores using Cosmos-Predict2–14B as a pixel-level world model for reward shaping, providing a fascinating case study of how (and why) straightforward world-model-based reward shaping sometimes falls short.

How It Works

• Proposes self-referential optimization: the VLA model generates its own reference trajectories and uses them for reward computation, eliminating external dependencies.
• Explores Cosmos-Predict2 as a pixel-level world model for generating reference trajectories based on language instructions.
• Finds that zero-shot Cosmos-Predict2–14B generation on LIBERO benchmark tasks produces visually plausible but functionally inaccurate trajectories.
• Develops SRPO as an alternative that bypasses the world model’s pixel-level limitations by operating in the policy’s own representation space.
• Evaluates on multiple manipulation benchmarks from the LIBERO suite.

Key Result

SRPO significantly outperforms prior VLA-RL methods including group-based optimization approaches. The paper’s analysis of Cosmos-Predict2 is particularly insightful: while the model generates visually compelling trajectories, the gap between “looks right” and “is functionally correct for reward shaping” remains significant, highlighting an important frontier for world model development.

Broader Impact

SRPO from Fudan and Tongji makes two important contributions. First, it offers a practical solution for improving VLA policies without expensive demonstrations. Second, its honest analysis of Cosmos-Predict2’s limitations as a reward model provides valuable guidance for the field: world models need to capture not just visual plausibility but functional correctness to serve as effective reward signals. This insight should inform the next generation of world foundation model training.

Why These Papers Matter — The Invisible Infrastructure of Physical AI

These five papers may not be the flashiest in the series, but they address the problems that determine whether Physical AI systems actually work in practice:

• AdapTok makes video tokenization adaptive and efficient, spending computational budget where it matters most.
• SceneScribe-1M provides the million-video dataset that world models need for joint geometric and semantic understanding.
• C-GenReg repurposes Cosmos-Transfer’s world knowledge for a completely new task — training-free 3D registration.
• Lifting Unlabeled Data turns the internet’s unlimited video supply into structured 3D training data.
• SRPO explores and honestly evaluates how world models can (and cannot yet) be used for policy optimization.

Together, they reveal a critical insight about where Physical AI stands today: the models are impressive, but the infrastructure, tokenizers, datasets, representations, and training strategies, is what determines whether those models translate into real-world capability.

Series Wrap-Up — 22 Papers, One Platform, Many Futures

Over five installments, we’ve explored 22 CVPR 2026 papers that cite NVIDIA Cosmos. Here’s what stands out:

Cosmos is becoming a platform, not just a model. Different teams use it in fundamentally different ways, as a backbone (VideoWorld 2), a baseline (PAI-Bench), a frozen foundation (PhyCo), a data generator (Real2Edit2Real), a geometry-to-image translator (C-GenReg), and a tokenizer (AdapTok). This diversity of usage is the hallmark of genuine platform adoption.

The field is honest about limitations. Papers like QUANTIPHY, SpatiaLQA, and SRPO don’t just showcase what Cosmos (releases before CVPR 2026) can do, they rigorously measure where it falls short. This intellectual honesty is what drives progress.

Physical AI is a global effort. These 22 papers come from 10 companies and 24+ universities across the US, China, Europe, Australia, and Israel. No single lab or country dominates; the field advances through distributed, international collaboration.

The gap between “looks right” and “is right” remains the central challenge. Video models generate visually stunning outputs, but physical correctness — the kind that lets a robot grasp an object or a car navigate safely — requires more than visual plausibility. Closing this gap is the next frontier.

Thank you for joining me on this journey through the Cosmos ecosystem at CVPR 2026. If any of these papers caught your eye, I encourage you to explore the full interactive report and reach out to the researchers behind them. The best collaborations often start with a conversation.

What is next?

If you’re interested in following along as I dive deeper into the world of AI and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Full blog about all 22 papers — Link

References

[1] Y. Li, C. Tian, R. Xia, N. Liao, W. Guo, J. Yan, H. Li, et al., “AdapTok: Learning Adaptive and Temporally Causal Video Tokenization in a 1D Latent Space,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2505.17011

[2] Y. Wang, K. Zheng, J. Wang, M. Chen, D. Novotny, et al., “SceneScribe-1M: A Large-Scale Video Dataset with Comprehensive Geometric and Semantic Annotations,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2604.07990

[3] Y. Haitman, A. Efraim, J. M. Francos, “C-GenReg: Training-Free 3D Point Cloud Registration by Multi-View-Consistent Geometry-to-Image Generation with Probabilistic Modalities Fusion,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2604.16680

[4] Y. Chen, Y. Zhang, H. Yu, J. He, Y. Wang, J. Huang, et al., “Lifting Unlabeled Internet-level Data for 3D Scene Understanding,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2604.01907

[5] S. Fei, S. Wang, L. Ji, A. Li, S. Zhang, L. Liu, J. Hou, et al., “SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2511.15605

Author: Paula Ramos, PhD

Welcome to Part 4 of our Cosmos Edition series. We’ve covered physical reasoning, embodied robotics, and driving simulation. Now we arrive at what might be the most exciting frontier: video generation as world modeling.

The idea is simple: if a model can generate accurate future video frames, it must have learned something meaningful about how the world work, object permanence, physics, cause and effect. But the gap between “visually plausible” and “physically correct” remains one of the biggest challenges in the field.

This installment highlights four papers that push the boundaries of what video generation can achieve, each using Cosmos models in different ways, as backbones, as baselines, or as frozen foundations to build upon.

The Strongest Cosmos Integration — VideoWorld 2 [1]

Paper Title: VideoWorld 2: Learning Transferable Knowledge from Real-world Videos
Paper Authors: Z. Ren, Y. Wei, X. Yu, G. Luo, Y. Zhao, B. Kang, et al.
Institution: ByteDance Seed, Beijing Jiaotong University

What if you could learn to fold paper, build towers, and manipulate objects, not from demonstrations, but from watching videos of humans doing it?

What It’s About

VideoWorld 2 presents the first investigation of learning transferable knowledge for complex, long-horizon tasks directly from raw real-world videos. At its core, it introduces a dynamics-enhanced Latent Dynamics Model (dLDM) that decouples action dynamics from visual appearance: a pretrained video diffusion model handles the visual side, while the dLDM learns compact latent codes that focus on task-related dynamics.

Why It Matters

This paper represents the deepest integration of Cosmos models in any CVPR 2026 paper. ByteDance doesn’t just cite Cosmos or benchmark against it — they use Cosmos AR 4B as their autoregressive transformer backbone and Cosmos DiT 2B as their appearance prior. It’s a validation that Cosmos models are robust enough to serve as core infrastructure for ambitious new architectures.

How It Works

• Uses the NVIDIA Cosmos AR 4B model as the autoregressive transformer, repurposing its next-token prediction capability to predict latent dynamics codes.
• Employs the Cosmos DiT 2B model as the appearance prior, leveraging its high-fidelity image-to-video capabilities to generate 93-frame (~5s at 16 fps) videos at 480px resolution.
• Decouples dynamics from appearance: the dLDM learns what happens (actions, state changes), while the DiT model handles how it looks (textures, lighting, details).
• Evaluates on Video-CraftBench — a new benchmark featuring real-world handcraft tasks like paper folding and block building.
• Also demonstrates cross-domain transfer: knowledge learned from Open-X robotics data improves performance on the CALVIN manipulation benchmark.

Key Result

VideoWorld 2 achieves up to 70% improvement in task success rate over baseline video generation models (including fine-tuned Cosmos AR and Cosmos DiT). On the CALVIN robotic manipulation benchmark, knowledge transferred from web videos substantially improves task performance, demonstrating genuine cross-domain generalization. The Cosmos backbone provides the visual generation quality, while the dLDM adds the dynamics understanding that raw generation models lack.

Broader Impact

VideoWorld 2 from ByteDance is a landmark paper for the Cosmos ecosystem. It demonstrates that Cosmos models can serve as building blocks for systems that go beyond video generation into genuine physical reasoning and task execution. The decoupling of dynamics from appearance is a powerful architectural insight that will likely influence future work in world modeling, robotics, and embodied AI.

Multi-View Magic — PlenopticDreamer [2]

Paper Title: Plenoptic Video Generation
Paper Authors: X. Fu, S. Tang, M. Shi, X. Liu, J. Gu, M.-Y. Liu, D. Lin, et al.
Institution: NVIDIA, The Chinese University of Hong Kong, Georgia Tech

Single-view video generation has made incredible progress. But what about generating consistent video from multiple camera viewpoints simultaneously?

What It’s About

PlenopticDreamer is a framework for camera-controlled generative video re-rendering that maintains spatio-temporal coherence across multiple views. While existing methods like ReCamMaster work well for single-view re-rendering, they struggle with multi-view consistency. PlenopticDreamer solves this by synchronizing generative hallucinations across views using spatio-temporal memory and camera-guided video retrieval.

Why It Matters

Multi-view consistency is critical for applications like robotic manipulation (where multiple cameras observe the workspace), autonomous driving (with surround-view cameras), and VR/AR content creation. If generated views are inconsistent, an object appears in one view but not another, the system is unreliable. PlenopticDreamer addresses this fundamental challenge.

How It Works

• Adopts Cosmos-Predict2.5–2B as the backbone model for video generation.
• Trains a multi-in-single-out video-conditioned model in an autoregressive manner: given previously generated views, the model generates the next view while maintaining consistency.
• Uses a camera-guided video retrieval strategy that selects the most relevant previously generated views as conditional inputs.
• Introduces progressive context-scaling during training to handle an increasing number of views.
• Evaluates on a new Basic Benchmark and the AgiBot robotic manipulation benchmark, demonstrating view transformations including third-person to third-person and head-view to gripper-view.

Key Result

PlenopticDreamer achieves state-of-the-art video re-rendering across all benchmarks, delivering superior view synchronization, high-fidelity visuals, accurate camera control, and diverse view transformations. On the Basic Benchmark, it consistently outperforms all baselines, including ReCamMaster retrained on Cosmos-Predict2.5, in view synchronization metrics across 3, 6, and 9 shots.

Broader Impact

PlenopticDreamer extends Cosmos from single-view video generation into the multi-view domain, which is where most real-world applications actually operate. The work demonstrates that Cosmos-Predict2.5 is a strong enough backbone to support multi-view consistency when combined with the right architectural innovations. This has direct implications for robotics, simulation, and immersive content creation.

One Token Per Frame — DeltaWorld [3]

Paper Title: A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens
Paper Authors: T. Kerssies, G. Berton, J. He, Q. Yu, W. Ma, et al.
Institution: Eindhoven University of Technology, Johns Hopkins University, Amazon

Existing world models represent each video frame with hundreds of spatial tokens. What if you could represent the change between frames with just one?

What It’s About

DeltaWorld introduces DeltaTok, a tokenizer that encodes the visual feature difference between consecutive frames into a single continuous “delta” token. Instead of representing video as a 3D spatio-temporal volume (which is computationally expensive), DeltaWorld reduces it to a 1D temporal sequence of delta tokens, enabling a small generative model to efficiently predict multiple diverse futures.

Why It Matters

Generative world models are powerful but computationally expensive. Cosmos-12B, for example, requires many forward passes and hundreds of spatial tokens per frame. DeltaWorld shows that by encoding only what changes between frames, rather than the full frame, you can build a much smaller model that still generates diverse, plausible futures. This efficiency gain could make world models practical for real-time applications.

How It Works

• DeltaTok encodes the feature difference (delta) between consecutive video frames using a vision foundation model (DINOv2), producing one continuous token per frame transition.
• DeltaWorld, the generative world model, operates on these delta tokens using a lightweight autoregressive transformer.
• Multiple future predictions can be generated in a single forward pass by sampling different delta token sequences.
• Benchmarked directly against Cosmos-12B on Cityscapes and KITTI driving datasets.
• Evaluates with task-specific heads (segmentation, depth) applied to the generated future features.

Key Result

DeltaWorld achieves competitive performance to Cosmos-12B on future prediction tasks while being dramatically more efficient, requiring a smaller model, fewer tokens, and the ability to generate multiple futures in a single forward pass. On Cityscapes, DeltaWorld maintains better fine-grained details (bicycle wheels, signposts, pedestrians) than Cosmos in short-horizon predictions.

Broader Impact

DeltaWorld from Amazon, JHU, and Eindhoven challenges the assumption that bigger is always better for world models. By rethinking the representation, encoding changes rather than states, they achieve competitive results at a fraction of the cost. This efficiency breakthrough could accelerate the deployment of world models in real-time systems like autonomous driving, where computational budgets are tight and multiple future scenarios need to be considered simultaneously.

Physics Meets Video Diffusion — PhyCo [4]

Paper Title: PhyCo: Learning Controllable Physical Priors for Generative Motion
Paper Authors: S. Narayanan, Z. Jiang, S. Narasimhan, et al.
Institution: Carnegie Mellon University, NEC Labs America

Video diffusion models create stunning visuals, but objects still float, collisions look wrong, and materials don’t behave realistically. What if we could inject actual physics into the generation process?

What It’s About

PhyCo introduces controllable, physically grounded control into video generation. It integrates three components: a large-scale dataset of photorealistic simulation videos where physical properties (friction, restitution, deformation, force) are systematically varied; physics-supervised fine-tuning of a Cosmos-Predict2–2B backbone using a ControlNet architecture; and VLM-guided reward optimization for physics-aware feedback.

Why It Matters

This paper is one of the strongest demonstrations of building directly on top of Cosmos. Rather than just benchmarking against it, PhyCo freezes the Cosmos-Predict2–2B backbone weights and fine-tunes ControlNet layers on top, using physical property maps as conditioning signals. The result is a model that inherits Cosmos’s video generation quality while adding genuine physical controllability.

How It Works

• Creates over 100K photorealistic simulation videos with systematically varied physical properties (friction, elasticity, deformation, force magnitude and direction).
• Fine-tunes ControlNet layers on a frozen Cosmos-Predict2–2B backbone, conditioning on pixel-aligned physical property maps.
• Each physical property is tokenized using the Cosmos Tokenizer, producing embeddings that condition the denoising process.
• Adds VLM-guided reward optimization: a fine-tuned VLM evaluates generated videos with physics-specific queries and provides differentiable feedback.
• Follows the same noise scheduling and temporal supervision as the Cosmos World Foundation Model training procedure.

Key Result

PhyCo significantly improves physical realism on the Physics-IQ benchmark compared to strong baselines. Human studies confirm clearer and more faithful control over physical attributes, users can specify friction, bounciness, and deformation, and the generated videos respond appropriately. The results demonstrate a scalable path toward physically consistent, controllable generative models.

Broader Impact

PhyCo from CMU and NEC Labs America is a blueprint for how to add domain-specific capabilities to foundation models without destroying their general capabilities. By keeping the Cosmos backbone frozen and training only the ControlNet layers, they preserve the pretrained dynamics while adding physics-aware control. This approach, freeze the foundation, fine-tune the adapter, is likely to become a standard pattern for building on top of world foundation models.

Why These Papers Matter — From Generating Videos to Understanding Worlds

These four papers represent different philosophies for how video generation connects to world understanding:

· VideoWorld 2 goes deepest, using Cosmos AR and DiT as the core backbone for learning task dynamics from raw video.

· PlenopticDreamer extends Cosmos into multi-view territory, solving the consistency problem that separates toy demos from real-world applications.

· DeltaWorld challenges the status quo, showing that efficient representations can match heavyweight models at a fraction of the cost.

· PhyCo adds physical grounding to Cosmos’s generation capabilities, freezing the backbone and fine-tuning physics-aware control layers.

Together, they suggest that Cosmos is evolving from a standalone model into a platform — a foundation that different teams build upon, extend, and challenge in complementary ways. ByteDance uses it as a backbone. NVIDIA extends it to multi-view. Amazon competes with it on efficiency. CMU adds physics on top of it. This ecosystem dynamic is exactly what a foundation model should enable.

Stay tuned for Part 5, where we close the series with papers on video tokenization, 3D scene understanding, and the data infrastructure that makes it all possible.

What is next?

If you’re interested in following along as I dive deeper into the world of AI and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Full blog about all 22 papers — Link

References

[1] Z. Ren, Y. Wei, X. Yu, G. Luo, Y. Zhao, B. Kang, et al., “VideoWorld 2: Learning Transferable Knowledge from Real-world Videos,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2602.10102

[2] X. Fu, S. Tang, M. Shi, X. Liu, J. Gu, M.-Y. Liu, D. Lin, et al., “Plenoptic Video Generation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2601.05239

[3] T. Kerssies, G. Berton, J. He, Q. Yu, W. Ma, et al., “A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2604.04913

[4] S. Narayanan, Z. Jiang, S. Narasimhan, et al., “PhyCo: Learning Controllable Physical Priors for Generative Motion,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2604.28169

Author: Paula Ramos, PhD

Welcome to Part 3 of our Cosmos Edition series. After exploring physical reasoning benchmarks and embodied robotics, we’re hitting the road, literally.

Autonomous driving is one of the most demanding applications of Physical AI. A self-driving system must predict how traffic evolves, handle rare safety-critical scenarios, and plan actions that are both safe and efficient. The challenge? These rare but critical scenarios are exactly the ones that are hardest to find in real-world driving data.

This is where world models shine. By learning the dynamics of driving environments, they can generate realistic simulations of scenarios that rarely occur in nature, aggressive lane changes, sudden braking, pedestrians appearing from blind spots. At CVPR 2026, four papers push this frontier forward, all referencing Cosmos models as part of their driving simulation toolkit.

Let’s take a ride through this research.

Training Better Drivers with Simulation at Scale — SimScale [1]

Paper Title: SimScale: Learning to Drive via Real-World Simulation at Scale
Paper Authors: H. Tian, T. Li, H. Liu, J. Yang, Y. Qiu, G. Li, J. Wang, et al.
Institution: OpenDriveLab at The University of Hong Kong, Xiaomi EV, Chinese Academy of Sciences.

What if instead of driving millions of miles, we could simulate millions of unseen scenarios from existing driving logs, and train better drivers in the process?

What It’s About

SimScale introduces a scalable simulation framework that synthesizes massive unseen driving states from existing real-world logs. By perturbing ego trajectories and using neural rendering with a reactive environment, it generates high-fidelity multi-view observations of scenarios that never actually happened, including safety-critical ones that are underrepresented in real data.

Why It Matters

Autonomous driving algorithms need exposure to rare events, aggressive lane switching, sudden obstacles, near-misses, to handle them safely. But these events are, by definition, uncommon in real-world data collected by human drivers. SimScale addresses this data diversity problem by generating realistic training scenarios at scale, without needing to put anyone in danger.

How It Works

• Takes existing driving logs and perturbs the ego vehicle’s trajectory to create new driving states.
• Uses advanced neural rendering to synthesize high-fidelity multi-view camera observations for each perturbed state.
• Generates pseudo-expert trajectories to provide action supervision for the new simulated states.
• Trains driving policies on a combination of real and simulated data with a co-training strategy.
• References Cosmos-Transfer1 as part of the landscape of neural rendering and controllable generation models for driving simulation.

Key Result

SimScale achieves up to +8.6 EPDMS improvement on the challenging navhard benchmark and +2.9 on navtest. Crucially, policy improvement scales smoothly by increasing simulation data alone, even without additional real-world data. This scaling property is what makes the approach truly practical.

Broader Impact

SimScale validates a key hypothesis: if simulation is realistic enough, we can scale driving policy training without scaling real-world data collection. This has massive implications for the cost and safety of autonomous driving development. The work from Xiaomi EV and the University of Hong Kong also signals growing international investment in simulation-driven autonomous driving, with Cosmos-Transfer1 recognized as part of the toolkit.

Editing Driving Scenes Like a Film Director — HorizonForge [2]

Paper Title: HorizonForge: Driving Scene Editing with Any Trajectories and Any Vehicles
Paper Authors: Y. Wang, F. Pittaluga, Z. Tasneem, C. You, et al.
Institution: Stony Brook University, NEC Labs America

What if you could take a normal driving video and edit it, inserting new vehicles, changing trajectories, and creating new scenarios, while maintaining photorealism and physical consistency?

What It’s About

HorizonForge is a unified framework for controllable driving scene editing. It reconstructs driving scenes as editable Gaussian Splats and Meshes, enabling fine-grained 3D manipulation and language-driven vehicle insertion. The edited scenes are rendered through a noise-aware video diffusion process that enforces spatial and temporal consistency, producing diverse variations in a single feed-forward pass.

Why It Matters

Testing autonomous driving algorithms requires diverse scenarios, but recording every possible traffic situation is impractical. HorizonForge enables researchers and engineers to take existing driving footage and edit it, inserting vehicles, changing trajectories, modifying lightingto create targeted test cases. This is more efficient than full simulation and more realistic than synthetic generation from scratch.

How It Works

• Reconstructs the original driving scene using Gaussian Splatting for the background and meshes for dynamic objects.
• Supports two types of editing: ego-level (changing the ego vehicle’s trajectory) and agent-level (inserting, removing, or modifying other vehicles).
• Uses a noise-aware video diffusion process to render edited scenes with spatial and temporal consistency.
• Introduces HorizonSuite, a benchmark spanning ego- and agent-level editing tasks.
• Cites Cosmos-Transfer1 as part of the broader ecosystem of controllable generation models for driving.

Key Result

HorizonForge achieves an 83.4% user-preference gain and a 25.19% FID improvement over the second-best state-of-the-art method. The framework produces photorealistic results in a single feed-forward pass without per-trajectory optimization, making it practical for large-scale scenario generation.

Broader Impact

HorizonForge democratizes driving scenario creation. Instead of needing massive simulation infrastructure, researchers can edit real driving footage to create targeted test cases. This approach from Stony Brook and NEC Labs America complements full-simulation approaches like Cosmos-Transfer by offering a lighter-weight, edit-based alternative that preserves the photorealism of real data.

When Planning Meets Prediction — DriveLaW [3]

Paper Title: DriveLaW: Unifying Planning and Video Generation in a Latent Driving World
Paper Authors: T. Xia, Y. Li, L. Zhou, J. Yao, K. Xiong, H. Sun, et al.
Institution: Huazhong University of Science and Technology

Driving world models can predict what happens next. Planners can decide what to do. But what if they could work together in the same latent space?

What It’s About

DriveLaW proposes a novel paradigm that unifies video generation and motion planning in a shared latent space. Instead of having a world model generate predictions that are then fed to a separate planner, DriveLaW directly injects latent representations from its video generator into the planner, ensuring inherent consistency between future prediction and trajectory planning.

Why It Matters

Current autonomous driving architectures keep world prediction and planning as decoupled processes, the world model imagines futures, and the planner picks actions, but the two don’t share a common understanding. DriveLaW argues that this decoupling is a fundamental limitation: the planner should reason about the same latent world the generator imagines, not a separate abstraction of it.

How It Works

• Learns generalizable driving representations through a video generation backbone that captures temporal continuity and dynamics.
• Directly injects latent video representations into the planning module, creating a unified latent driving world.
• References Cosmos as one of the key world models — alongside Genie — that learn real-world structure from large-scale video generation.
• Also cites Cosmos-Drive-Dreams as a large-scale synthetic data pipeline for driving, positioning DriveLaW in the context of recent industry efforts.

Key Result

DriveLaW demonstrates that unified latent planning outperforms decoupled architectures on standard driving benchmarks, producing trajectories that are more consistent with predicted scene evolution. The unification eliminates the information loss that occurs when translating between separate prediction and planning representations.

Broader Impact

DriveLaW from Huazhong University of Science and Technology points toward a future where autonomous driving systems don’t just predict and then plan, they think in a single, coherent world representation. This mirrors the broader vision of world foundation models like Cosmos: learning a unified representation of how the world works that can serve multiple downstream tasks simultaneously.

Less Compute, More Physics — MAD Driving World Models [4]

Paper Title: MAD: Motion Appearance Decoupling for Efficient Driving World Models (Website)
Paper Authors: A. Rahimi, V. Gerard, E. Zablocki, M. Cord, Alexandre Alahi.
Institution: Sorbonne Université, Valeo AI, EPFL

Models like Cosmos-Predict are trained from scratch on massive datasets with enormous compute budgets. But what if you could build a competitive driving world model for a fraction of the cost?

What It’s About

MAD (Motion Appearance Decoupling) introduces an efficient approach to building driving world models by separating the problem into two stages: a motion forecaster that predicts how the scene evolves dynamically, and an appearance synthesizer that renders the final video. This decoupling lets each component focus on what it does best, dramatically reducing the compute needed to achieve competitive results.

Why It Matters

State-of-the-art driving world models like Cosmos-Predict are powerful but resource-intensive, trained from scratch on massive proprietary datasets requiring compute budgets that are out of reach for most academic labs. MAD shows that by thinking carefully about architecture design, you can build a model that achieves comparable generation quality using only a fraction of the compute and data.

How It Works

• Decouples driving video generation into motion prediction (how things move) and appearance synthesis (how things look).
• The motion forecaster operates on a compact intermediate representation, avoiding the need to predict full-resolution pixels.
• The appearance synthesizer handles the final rendering step, leveraging the motion forecast as conditioning.
• Builds on the LTX backbone and achieves generation quality comparable to Cosmos-Predict2 while using significantly less compute.
• Supports conditioning on text, ego-motion, and 3D bounding boxes for controllable generation.

Key Result

MAD-LTX outperforms all previous open driving world models and achieves generation quality comparable to Cosmos-Predict2, while using a fraction of the compute and data budget. The key insight: by decoupling motion from appearance, the model can focus its limited capacity on learning the dynamics that matter for driving, rather than spending it on rendering textures and lighting.

Broader Impact

MAD from Sorbonne and Valeo AI makes a powerful statement: you don’t need industrial-scale resources to build competitive driving world models. By thoughtfully decomposing the problem, academic labs and smaller companies can contribute meaningfully to the driving simulation space. The paper also serves as independent validation that Cosmos-Predict sets a high bar, being competitive with it is genuinely noteworthy.

Why These Papers Matter — Simulation as the New Training Ground

These four papers paint a vivid picture of how the autonomous driving field is evolving:

• SimScale shows that realistic simulation can replace real-world data collection at scale.
• HorizonForge enables targeted scenario editing — creating the exact test cases you need from existing footage.
• DriveLaW unifies prediction and planning in a shared latent space, eliminating the gap between imagining the future and deciding what to do.
• MAD democratizes driving world models, proving you don’t need massive compute to achieve state-of-the-art results.

What connects them all is a shared conviction: the future of autonomous driving runs through simulation, and world foundation models like Cosmos are becoming the infrastructure that makes it possible. Whether teams are benchmarking against Cosmos, building on top of it, or competing with it on a fraction of the budget, its presence across these papers confirms that world models have moved from research curiosity to engineering necessity.

Stay tuned for Part 4, where we explore how Cosmos is powering the next generation of video generation and world modeling, from paper crafts to plenoptic views.

What is next?

If you’re interested in following along as I dive deeper into the world of AI and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Full blog about all 22 papers — Link

References

[1] H. Tian, T. Li, H. Liu, J. Yang, Y. Qiu, G. Li, J. Wang, et al., “SimScale: Learning to Drive via Real-World Simulation at Scale,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2511.23369

[2] Y. Wang, F. Pittaluga, Z. Tasneem, C. You, et al., “HorizonForge: Driving Scene Editing with Any Trajectories and Any Vehicles,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2602.21333

[3] T. Xia, Y. Li, L. Zhou, J. Yao, K. Xiong, H. Sun, et al., “DriveLaW: Unifying Planning and Video Generation in a Latent Driving World,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2512.23421

[4] A. Rahimi, V. Gerard, E. Zablocki, M. Cord, et al., “MAD: Motion Appearance Decoupling for Efficient Driving World Models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2601.09452

Author: Paula Ramos, PhD

Welcome back to our Cosmos Edition of “The Best of CVPR.” In Part 1, we explored the benchmarks that are setting the standard for Physical AI reasoning. Now we step from evaluation into action.

This second installment highlights five CVPR 2026 papers focused on embodied intelligence, giving robots the ability to reason about their environment, plan multi-step tasks, recover from failures, and learn from synthetic data. What connects them all is the use of NVIDIA Cosmos models as reasoning backbones, evaluation baselines, or data generation engines. These papers show us what happens when world models meet real-world robotic systems.

Let’s see what’s happening at the frontier of robot learning.

Drawing the Plan Before Grasping — Action-Sketcher [1]

Paper Title: Action-Sketcher: From Reasoning to Action via Visual Sketches for Robotic Manipulation
Paper Authors: H. Tan, P. Co, Y. Xu, S. Rong, Y. Ji, C. Chi, X. Chen, et al.
Institution: Peking University, Intel, Chinese Academy of Sciences

What if a robot could sketch out its plan visually, drawing arrows, circles, and paths on the scene, before executing a single action?

What It’s About

Action-Sketcher introduces a novel approach to long-horizon robotic manipulation where the robot first generates a visual sketch of its intended action plan directly on the scene image, then uses that sketch to guide execution. Instead of keeping the plan hidden inside text embeddings, the system externalizes its reasoning as visual annotations, making the plan interpretable and debuggable.

Why It Matters

Long-horizon robotic manipulation, tasks that require multiple sequential actions in cluttered environments, is one of the hardest problems in embodied AI. Current Vision-Language-Action (VLA) policies typically rely on text-only cues and keep their plans latent, making it hard to diagnose failures. Action-Sketcher makes the reasoning visible, which improves both grounding and task decomposition.

How It Works

• Given a scene image and a language instruction, a VLM generates a visual sketch (arrows, regions, trajectories) overlaid on the RGB observation.
• The sketch serves as an intermediate representation that bridges high-level reasoning and low-level action.
• The system is evaluated against embodied baselines including Cosmos-Reason1–7B across five spatial reasoning benchmarks (BLINK, CV-Bench, EmbSpatial, RoboSpatial, RefSpatial-Bench).
• Supports close-loop interaction for long-horizon tasks, updating sketches based on new observations.

Key Result

Action-Sketcher outperforms both general VLMs and specialized embodied models (including Cosmos-Reason1–7B) on spatial reasoning benchmarks. The visual sketch representation provides stronger referential grounding than text-only plans, especially in cluttered or ambiguous scenes where language descriptions are insufficient.

Broader Impact

Action-Sketcher offers a compelling paradigm shift: instead of hoping that latent representations capture the right plan, let the model draw its intentions explicitly. This makes robotic systems more interpretable and easier to debug, a critical requirement for real-world deployment. It also sets a new baseline for how we evaluate embodied reasoning, with Cosmos-Reason1 serving as a key reference point.

Chaining Skills Like Building Blocks — RoboAgent [2]

Paper Title: RoboAgent: Chaining Basic Capabilities for Embodied Task Planning
Paper Authors: P. Xu, J. Zheng, Y. Mu
Institution: Peking University, XYZ Embodied AI

A robot can pick up an object, pour liquid, and open a drawer — but can it chain these skills together to make a cup of coffee?

What It’s About

RoboAgent tackles the problem of embodied task planning by decomposing complex tasks into chains of basic capabilities. Rather than training a single end-to-end policy for every possible task, the system learns to identify, sequence, and execute atomic skills in the right order to accomplish long-horizon goals.

Why It Matters

End-to-end VLA models struggle with compositional tasks, they can learn individual skills from demonstrations but often fail when those skills need to be combined in novel sequences. RoboAgent addresses this by treating task planning as a chaining problem, which scales more naturally to the combinatorial diversity of real-world tasks.

How It Works

• Decomposes tasks into a library of basic capabilities (grasp, place, pour, open, close).
• Uses a visual observation-based planner to select and sequence capabilities based on the current scene state.
• Cites Cosmos-Reason1 as part of the broader context of embodied reasoning models that inform their approach to grounding actions in physical understanding.
• Evaluates on multi-step manipulation scenarios requiring both spatial reasoning and temporal sequencing.

Key Result

RoboAgent demonstrates that capability chaining significantly improves success rates on long-horizon tasks compared to monolithic end-to-end approaches. The modular structure also enables better generalization to unseen task combinations, since new behaviors can emerge from novel sequences of already-learned skills.

Broader Impact

RoboAgent points toward a future where robots build complex behaviors from simple, reusable primitives, much like how we build software from libraries and APIs. This modular approach also makes it easier to diagnose failures (which skill failed?) and update the system (add a new skill without retraining everything). The work builds on the growing ecosystem of embodied reasoning models, where Cosmos-Reason1 is one of the key references.

Learning from Failure — When Robots Get a Second Chance [3]

Paper Title: Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols
Paper Authors: X. Zeng, X. Zhou, Y. Li, J. Shi, T. Li, L. Chen, L. Ren, et al.
Institution: Beihang University, Shanghai Jiao Tong University, Southern University of Science and Technology

Robots fail. A lot. But what if, instead of just trying again, they could diagnose what went wrong, correct their approach, and learn from the experience?

What It’s About

This paper introduces ViFailback, a system that enables VLA models to diagnose manipulation failures, generate corrective strategies, and accumulate failure-based knowledge over time. It uses visual symbols, structured visual representations of failure modes, to create an interpretable failure diagnosis framework that goes beyond binary success/fail signals.

Why It Matters

Current VLA models like π0.5 are impressive at following demonstrations, but they remain brittle when things go wrong. A missed grasp, a slipped object, or an unexpected collision typically leads to task failure with no recovery. ViFailback addresses this by building a failure diagnosis and correction loop on top of existing VLA policies, making robots more resilient in practice.

How It Works

• Monitors manipulation execution and detects failures using visual analysis of the scene state.
• Diagnoses the failure type (missed grasp, collision, wrong orientation) using visual symbols — structured annotations that categorize and localize the failure.
• Generates corrective actions based on the diagnosis.
• Benchmarks failure diagnosis capabilities against Cosmos-Reason1–7B and other models including GPT-4o, Gemini-2.5-Pro, RoboBrain2.0, and general VLMs.
• Integrates with the π0.5 VLA model for real-world manipulation experiments.

Key Result

ViFailback-8B significantly improves recovery rates in real-world manipulation tasks when integrated with π0.5. In failure diagnosis benchmarks, the system outperforms Cosmos-Reason1–7B (38.06% vs. 33.75% average) and approaches the performance of much larger closed-source models. The key insight: specialized failure-aware training matters more than raw model scale for this task.

Broader Impact

Failure recovery is one of the most underexplored aspects of robot learning. This paper makes a compelling case that we need to train not just policies but also critics and diagnosticians that understand why things go wrong. It also demonstrates that Cosmos-Reason1 provides a strong baseline for physical reasoning about failures, validating its relevance as infrastructure for embodied AI.

Imagining the Future in 3D — PointWorld [4]

Paper Title: PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation
Paper Authors: W. Huang, Y.-W. Chao, A. Mousavian, M.-Y. Liu, et al.
Institution: Stanford University, NVIDIA

Humans can glance at a scene and anticipate how it will change when they act on it. Can a robot do the same in 3D?

What It’s About

PointWorld is a large pre-trained 3D world model that unifies state and action in a shared 3D space as 3D point flows. Given one or a few RGB-D images and a sequence of robot action commands, PointWorld forecasts per-pixel displacements in 3D that respond to the given actions. By representing actions as 3D point flows instead of embodiment-specific spaces (like joint positions), the model generalizes across different robot platforms.

Why It Matters

Most world models operate in 2D image space or abstract latent spaces. PointWorld works directly in 3D, which means its predictions are physically grounded, they respect geometry, occlusion, and depth. This is critical for manipulation, where a robot needs to understand that pushing an object sideways will also change its appearance from the camera’s perspective.

How It Works

• Represents the world as 3D point clouds and learns to predict 3D point flows (displacements) conditioned on robot actions.
• Trained at scale on approximately 2 million trajectories and 500 hours of manipulation data across single-arm Franka and bimanual humanoid robots.
• Achieves real-time inference (0.1s), enabling integration into model-predictive control (MPC) frameworks.
• Cites the Cosmos World Foundation Model Platform as part of the broader landscape of world models that PointWorld extends into the 3D domain.

Key Result

A single pre-trained PointWorld checkpoint enables a real-world Franka robot to perform rigid-body pushing, deformable and articulated object manipulation, and tool use — without requiring any task-specific demonstrations. The model’s 3D representation enables cross-embodiment transfer: knowledge learned from a humanoid robot transfers to a single-arm setup.

Broader Impact

PointWorld represents a significant step toward universal robot manipulation through 3D world modeling. By operating in 3D point space rather than image space, it achieves a level of physical grounding that 2D approaches cannot match. The connection to the Cosmos platform is natural: both aim to learn generalizable representations of how the world works, with PointWorld focusing specifically on the 3D manipulation domain.

From Five Demos to Fifty — Real2Edit2Real [5]

Paper Title: Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface
Paper Authors: Y. Zhao, H. Fan, D. Chen, S. Chen, L. Chen, X. Li, et al.
Institution: Peking University, AgiBot

Collecting robotic demonstrations is expensive and tedious. What if you could collect just five and generate fifty more, all with realistic 3D variations?

What It’s About

Real2Edit2Real presents a framework for data augmentation in robot learning: it takes a small number of real-world demonstrations, reconstructs the 3D scene geometry, edits object positions and configurations in 3D, and then generates photorealistic multi-view manipulation videos using a video generation model fine-tuned on Cosmos-Predict-2B. The synthetic demonstrations are realistic enough to train effective manipulation policies.

Why It Matters

The bottleneck in robot learning isn’t algorithms — it’s data. Collecting diverse demonstrations for spatial generalization is time-consuming and doesn’t scale. Real2Edit2Real breaks this bottleneck by generating new demonstrations from 3D edits of existing ones, achieving 10–50x data efficiency improvements. This has direct implications for anyone trying to deploy robotic systems in new environments.

How It Works

• Reconstructs scene geometry from multi-view RGB observations using a metric-scale 3D reconstruction model.
• Performs 3D edits (object repositioning, orientation changes) on the reconstructed point cloud via an intuitive 3D control interface.
• Renders consistent depth maps from the edited scenes, which serve as conditioning signals.
• Uses a multi-conditional video generation model (based on Cosmos-Predict-2B) guided by depth, action, edge, and ray maps to synthesize photorealistic manipulation videos.
• Trains visuomotor policies on the combination of real and generated demonstrations.

Key Result

Policies trained on data generated from only 1–5 source demonstrations can match or outperform those trained on 50 real-world demonstrations. The framework achieves 10–50x improvement in data efficiency across four real-world manipulation tasks, with additional experiments demonstrating flexibility for height and texture editing.

Broader Impact

Real2Edit2Real demonstrates that Cosmos-Predict models can serve as practical data generation engines for robotics, not just research artifacts. By bridging 3D editability with 2D video synthesis, it offers a scalable path to train robots for new tasks and environments without the prohibitive cost of manual data collection. This is exactly the kind of downstream application that validates the Cosmos platform’s vision of being foundational infrastructure for Physical AI.

Why These Papers Matter — From Understanding to Doing

If Part 1 of this series was about asking “can AI understand physics?”, Part 2 answers a different question: “can AI act on that understanding?”

These five papers reveal the full loop of embodied intelligence:

• Action-Sketcher shows that making plans visible improves robotic reasoning.
• RoboAgent demonstrates that chaining simple skills enables complex behaviors.
• ViFailback proves that learning from failure is as important as learning from success.
• PointWorld takes world modeling into 3D, where physical manipulation actually happens.
• Real2Edit2Real uses Cosmos as a data engine, turning a handful of demonstrations into scalable training sets.

What’s exciting is how these papers complement each other. Imagine a system that plans with visual sketches, chains atomic skills, recovers from failures using physics-aware diagnosis, predicts outcomes in 3D, and generates its own training data. That’s not science fiction, it’s the research trajectory these papers collectively outline.

Stay tuned for Part 3, where we hit the road with Cosmos-powered driving world models.

What is next?

If you’re interested in following along as I dive deeper into the world of AI and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Full blog about all 22 papers — Link

References

[1] H. Tan, P. Co, Y. Xu, S. Rong, Y. Ji, C. Chi, X. Chen, et al., “Action-Sketcher: From Reasoning to Action via Visual Sketches for Robotic Manipulation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. https://arxiv.org/abs/2601.01618

[2] P. Xu, J. Zheng, and Y. Mu, “RoboAgent: Chaining Basic Capabilities for Embodied Task Planning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. https://arxiv.org/abs/2604.07774

[3] X. Zeng, X. Zhou, Y. Li, J. Shi, T. Li, L. Chen, L. Ren, et al., “Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. https://arxiv.org/abs/2512.02787

[4] W. Huang, Y.-W. Chao, A. Mousavian, M.-Y. Liu, et al., “PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. https://arxiv.org/abs/2601.03782

[5] Y. Zhao, H. Fan, D. Chen, S. Chen, L. Chen, X. Li, et al., “Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. https://arxiv.org/abs/2512.19402

Author: Paula Ramos, PhD

The first generation of Physical AI was about making things work, robots that grasp, cars that drive, models that generate video. But before we build further, we need to ask a harder question: do these systems actually understand physics?

Four papers at CVPR 2026 say: let’s find out. They introduce benchmarks that go beyond visual quality, probing forces, causality, spatial logic, and quantitative estimation. The results reveal exactly where the frontier stands, and why Cosmos is the reference point the community builds on

Let’s dive in.

Measuring the Full Spectrum of Physical AI — PAI-Bench [1]

Paper Title: PAI-Bench: A Comprehensive Benchmark For Physical AI
Paper Authors: F. Zhou, J. Huang, J. Li, D. Ramanan, H. Shi
Institution: Carnegie Mellon University

What if we had a single benchmark that could test everything a Physical AI model needs to do, perceive, predict, and reason about real-world dynamics? PAI-Bench does exactly that, and the results are humbling.

What It’s About

PAI-Bench is a unified benchmark that evaluates both video generative models and multimodal large language models on their ability to handle the perceptual and predictive demands of Physical AI. It includes 2,808 real-world test cases with task-aligned metrics designed to capture physical plausibility and domain-specific reasoning across video generation, conditional video generation, and video understanding.

Why It Matters

The Physical AI space is fragmented: some benchmarks test video quality, others test reasoning, and few test both in a unified way. PAI-Bench fills this gap by providing a comprehensive evaluation framework. Crucially, it evaluates models like Cosmos-Predict2 (both 2B and 14B), Cosmos-Transfer, and Cosmos-Reason alongside other state-of-the-art systems, giving us the most complete picture yet of where these models excel and where they fall short.

How It Works

• Covers three evaluation axes: video generation quality, conditional generation fidelity (given structural inputs like depth, edges, segmentation), and video understanding (physical reasoning questions).
• Tests 20+ models including Cosmos-Predict2–2B, Cosmos-Predict2–14B, Cosmos-Predict2.5–2B, and Cosmos-Transfer1–7B alongside competitors like Wan2.1, CogVideoX, and MAGI-1.
• Uses task-aligned metrics that capture physical plausibility — not just visual fidelity — including blur consistency, edge accuracy, depth estimation, and mask IoU.
• Pairs quantitative metrics with human evaluation for conditional generation quality and diversity.

Key Result

Video generative models, despite producing visually impressive outputs, often struggle to maintain physically coherent dynamics. Multimodal language models show limited performance in forecasting and causal interpretation. Among the Cosmos models, Cosmos-Predict2.5–2B achieves competitive scores across the board, particularly in conditional generation tasks, but even the best models reveal significant room for improvement in physical reasoning.

Broader Impact

PAI-Bench establishes a realistic and comprehensive foundation for evaluating Physical AI. By systematically exposing where current systems fail, particularly in maintaining physical coherence over time, it provides a clear roadmap for the next generation of world models. For anyone building on Cosmos or competing with it, this benchmark is the new measuring stick.

Building Physics-Aware Judges for Embodied AI — PhyCritic [2]

Paper Title: PhyCritic: Multimodal Critic Models for Physical AI
Paper Authors: T. Xiong, S. Wang, G. Liu, Y. Dong, M. Li, H. Huang, et al.
Institutions: Tsinghua University, Shanghai Jiao Town University, Shanghai AI Laboratory

We have models that can see, reason, and act, but who judges whether their physical reasoning is actually correct? PhyCritic introduces the first critic model designed specifically for Physical AI.

What It’s About

PhyCritic is a specialized multimodal critic model designed to evaluate physical AI tasks involving perception, causal reasoning, and action-centric planning. Rather than asking “did the robot complete the task?”, it asks deeper questions: “was the reasoning physically valid? Did the model understand causality? Did it ground its decisions in visual evidence?” The paper also introduces PhyCritic-Bench, a novel benchmark built to rigorously evaluate multimodal judge models in physical AI scenarios using diverse embodied datasets.

Why It Matters

As embodied AI systems grow more capable, we need equally capable evaluation systems. Existing multimodal critic models typically lack physics awareness and fail to detect physically impossible reasoning. PhyCritic addresses this by employing a two-stage training pipeline (physical skill warmup and self-referential critic finetuning), creating a judge that can deliver stable, physically correct judgments based on underlying causal validity.

How It Works

• Builds a high-quality physical critic dataset spanning diverse embodied domains including RoboVQA, BridgeData V2, HoloAssist, and AgiBot World.
• Uses self-referential physical grounding: the model generates its own internal prediction as a reference before evaluating candidate responses.
• Collects model outputs from multiple multimodal models to create a benchmark for physical reasoning.
• The training pipeline includes Group Relative Policy Optimization (GRPO) to enhance basic perception and reasoning.

Key Result

PhyCritic outperforms existing open-source 7B and 8B baselines on both physical and general reward benchmarks. These improvements span multiple evaluations, including Cosmos-Reason1, CV-Bench, EgoPlan-Bench2, and VL-RewardBench, demonstrating that self-referential physical grounding significantly benefits critic models.

Broader Impact

PhyCritic signals a new direction for the field: as Physical AI models grow in capability, we need specialized evaluation infrastructure to keep pace. The idea that a critic model should understand physics — not just language and vision — opens the door for more reliable training pipelines and safer robotic systems. It also validates the need for foundational resources specifically designed for physically grounded systems.

Does Your AI Really Understand the Physical World? — QUANTIPHY [3]

Paper Title: QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models
Paper Authors: L. Puyin, T. Xiang, E. Mao, S. Wei, X. Chen, et al.
Institution: Stanford University

Understanding the physical world is essential for generalist AI agents. But do today’s best vision models actually know how heavy something is, how fast it’s moving, or whether it will fall?

What It’s About

QUANTIPHY is a benchmark designed to evaluate whether vision-language models can reason quantitatively about physical properties, not just describe what they see, but estimate real-world measurements like mass, velocity, force, and distance. It goes beyond qualitative reasoning (“this is heavy”) to demand quantitative precision (“this weighs approximately 2 kg”).

Why It Matters

Most physical reasoning benchmarks test whether models can describe what’s happening in a scene. But real-world physical AI, robots, self-driving cars, manufacturing systems, needs quantitative understanding. A robot needs to know not just that a box is heavy, but how much force to apply. QUANTIPHY exposes a critical gap: even the best VLMs struggle with quantitative physical reasoning.

How It Works

• Presents real-world scenarios requiring estimation of physical quantities (mass, speed, distance, force, acceleration).
• Evaluates state-of-the-art VLMs including GPT-4o, Gemini, Claude, and domain-specific models.
• Tests both zero-shot and chain-of-thought reasoning strategies.
• Provides fine-grained error analysis showing where each model systematically fails.

Key Result

State-of-the-art VLMs perform surprisingly poorly on quantitative physical reasoning, even models specifically trained for physical AI. The benchmark reveals that current models can often describe physical scenarios qualitatively but lack the ability to produce accurate numerical estimates, suggesting that physical reasoning training needs to incorporate more quantitative grounding.

Broader Impact

QUANTIPHY highlights a fundamental gap that the field must address: qualitative understanding of physics is not enough for real-world deployment. As world models evolve, this benchmark provides a critical test of whether they’re learning genuine physical intuition or just pattern-matching on visual cues.

When Seeing Isn’t Enough — Spatial Logical Reasoning for VLMs — SpatiaLQA [4]

Paper Title: SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models
Paper Authors: Y. Xie, X. Zhang, Y. Shan, H. Zhu, R. Tang, R. Wei, et al.
Institutions: Zhejiang University, The University of Sydney, ManyCore

VLMs can answer questions about images and even reason logically , but can they combine spatial understanding with multi-step logical reasoning in complex real-world environments?

What It’s About

SpatiaLQA introduces a benchmark that evaluates spatial logical reasoning, the ability to understand not just where objects are in a scene, but how spatial relationships constrain the logical sequence of actions needed to complete a task. Think of it as asking: “In what order do you need to remove these stacked objects to reach the one at the bottom?”

Why It Matters

Real-world tasks, from warehouse logistics to kitchen robotics, require both spatial awareness and logical planning. A robot stacking boxes needs to understand that you can’t remove the bottom box without first moving the ones on top. Current VLMs handle common VQA and logical reasoning well in isolation but fail when both capabilities must work together.

How It Works

• Defines spatial logical reasoning as a composite capability requiring spatial understanding + multi-step logical dependencies.
• Provides scenarios with stacked, nested, and occluded objects where the sequence of actions matters.
• Evaluates 20+ VLMs including Cosmos-Reason1–7B, GPT-4o, Qwen3-VL, InternVL3.5, and others.
• Measures both spatial understanding accuracy and logical consistency of multi-step plans.

Key Result

Cosmos-Reason1–7B achieves 48.2% accuracy on single-step spatial reasoning and 56.1% overall, placing it competitively among 7B-class models. However, all models, including much larger ones, show dramatic performance drops when multi-step logical reasoning is required on top of spatial understanding, with scores falling below 20% on the hardest compositional tasks.

Broader Impact

SpatiaLQA reveals that spatial understanding and logical reasoning remain poorly integrated in current VLMs. This has direct implications for embodied AI, where robots must plan physically feasible action sequences in cluttered environments. The benchmark provides a clear signal: the next generation of physical reasoning models needs to combine spatial perception with structured planning.

Why These Papers Matter — Setting the Standard for Physical AI

These four CVPR 2026 papers share a common mission: defining what it means for AI to truly understand the physical world. Together, they paint a nuanced picture:

• PAI-Bench gives us the comprehensive evaluation framework, testing perception, prediction, and generation in one place.
• PhyCritic builds the judge, a physics-aware critic that can tell us whether reasoning is genuinely grounded in physical understanding.
• QUANTIPHY pushes beyond qualitative understanding to demand quantitative precision, the kind real-world robots actually need.
• SpatiaLQA combines spatial and logical reasoning, exposing a critical gap in how models plan actions in physical spaces.

What stands out is the honesty of these evaluations. All four papers show that even the most advanced models have significant room for improvement. This isn’t a weakness; it’s a sign of a maturing field that’s willing to measure itself rigorously before claiming progress.

The message is clear: Physical AI is not just about generating beautiful videos or answering questions about images. It’s about building systems that understand forces, predict consequences, and plan actions in the real world. These benchmarks set the bar, and the race to clear it is just beginning.

Stay tuned for Part 2, where we explore how Cosmos is being used to teach robots to think, plan, and act in the physical world.

What is next?

If you’re interested in following along as I dive deeper into the world of AI and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Full blog about all 22 papers — Link

References

[1] F. Zhou, J. Huang, J. Li, D. Ramanan, and H. Shi, “PAI-Bench: A Comprehensive Benchmark For Physical AI,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2512.01989

[2] T. Xiong, S. Wang, G. Liu, Y. Dong, M. Li, H. Huang, et al., “PhyCritic: Multimodal Critic Models for Physical AI,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2602.11124

[3] L. Puyin, T. Xiang, E. Mao, S. Wei, X. Chen, et al., “QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2512.19526

[4] Y. Xie, X. Zhang, Y. Shan, H. Zhu, R. Tang, R. Wei, et al., “SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026. Temporal link: https://arxiv.org/abs/2602.20901

Author: Paula Ramos, PhD

I just spent some weeks reading, writing about, and living inside 22 CVPR 2026 papers that cite NVIDIA Cosmos. Five blog posts. Four thousand accepted papers at the conference. One question that kept coming back to me: Where is all of this actually going?

If you’ve been following my “Best of CVPR” series, you know I don’t write these to summarize papers, I write them to find the signal in the noise. CVPR is overwhelming. The massive volumen of research makes it hard to see the patterns, the gaps, the inflection points. That’s what this series was about: not just “what’s new,” but “what’s changing.”

My intention is to put it all together, the threads that connect 22 papers, five blog posts, and the biggest open questions for the next year. If you didn’t have time to read the full series, start here. And if you did read it all, this is where we step back and see the map.

How I Found These 22 Papers: The Process Behind the Series

This series wasn’t just a random selection of papers. It was the result of a deliberate, week-long deep dive into the CVPR 2026 proceedings, specifically looking for research that cites NVIDIA Cosmos. My goal was to understand how the community is adopting, extending, and even challenging the world model foundations we’ve built.

The process was simple but intense: I started with a pool of nearly 100 papers that mentioned “Cosmos” or “world models” in the context of physical AI. From there, I filtered for papers that provided either a significant technical extension (like new tokenization methods) or a rigorous benchmark of physical reasoning. What remained were the 22 papers that I believe define the state of the art in Physical AI today.

A Quick Orientation of Five Chapters

Before we zoom out, here’s where we’ve been:

Part 1 — Can AI Understand Physics? Four benchmarks (PAI-Bench [1], PhyCritic [2], QUANTIPHY [3], SpatiaLQA [4]) that test whether world models genuinely understand forces, causality, spatial reasoning, and quantitative physics. Spoiler: even the best models have a long way to go.

Part 2 — Teaching Robots to Think, Plan, and Recover Five papers on embodied intelligence — robots that sketch their plans visually (Action-Sketcher [5]), chain simple skills into complex behaviors (RoboAgent [6]), diagnose and learn from failures (ViFailback [7]), predict outcomes in 3D (PointWorld [8]), and generate their own training data from five demonstrations (Real2Edit2Real [9]).

Part 3 — The Road Ahead Four driving world models that show simulation is becoming the primary training ground for autonomous vehicles — from generating millions of scenarios (SimScale [10]) to editing scenes like a film director (HorizonForge [11]) to unifying planning and prediction in a single latent space (DriveLaW [12]) to efficient motion-appearance decoupling (MAD [13]).

Part 4 — Seeing the Future Four papers pushing video generation from “looks right” toward “is right” — learning task dynamics from raw video (VideoWorld 2 [14]), generating multi-view consistency (PlenopticDreamer [15]), one-token-per-frame efficiency (DeltaWorld [16]), and injecting real physics into diffusion models (PhyCo [17]).

Part 5 — The Invisible Infrastructure Five papers on the foundation layer — adaptive video tokenization (AdapTok [18]), million-scale annotated datasets (SceneScribe-1M [19]), using world models for 3D registration without training (C-GenReg [20]), turning unlabeled internet video into 3D training data (Lifting Unlabeled Data [21]), and the honest limits of world models as reward signals (SRPO [22]).

Three Big Themes That Changed How I Think About Physical AI

1. We’re moving from “can it see?” to “can it do?”

The first generation of deep learning was about perception: can the model recognize a cat, a car, a tumor? The second wave was about generation: can it create images, write text, produce video? CVPR 2026 is the beginning of a third wave: can it act?

Papers like PointWorld [8], RoboAgent [6], and Action-Sketcher [5] aren’t asking whether a model can generate a plausible video of a robot grasping a cup. They’re asking whether that same understanding of physics can drive a real robot arm to actually pick it up. ViFailback [7] goes further: when the robot drops the cup, can it figure out why and try a different approach?

This is the shift that matters. Video generation was the proof of concept. Physical AI: robots, autonomous vehicles, industrial automation, is the deployment target. And the gap between generating a beautiful video and producing a physically correct action is where the hardest research problems live right now.

2. World models are becoming platforms, not papers

A year ago, Cosmos was a model you benchmarked against. At CVPR 2026, it’s infrastructure people build on, and that shift is the biggest story of this conference.

ByteDance uses Cosmos AR 4B as the backbone for VideoWorld 2 [14], achieving 70% improvement in task success by decoupling dynamics from appearance on top of Cosmos's visual generation. CMU freezes Cosmos-Predict2 and fine-tunes physics-aware ControlNet layers on top with PhyCo [17], adding physical controllability without sacrificing the foundation’s quality. NVIDIA extends it to multi-view plenoptic generation with PlenopticDreamer [15]. Ben-Gurion University repurposes Cosmos-Transfer for 3D point cloud registration with C-GenReg [20], a task that has nothing to do with video generation. Even teams exploring alternative architectures, like Amazon’s DeltaWorld [16], use Cosmos-12B as the benchmark to validate against, which tells us what the standard is.

This is what platform adoption looks like. Not everyone using the same model the same way, but different teams finding different value in the same foundation. Some extend it. Some build on it. Some use it for purposes its creators never imagined. That diversity , 22 papers from 10 companies and 24+ universities, is the strongest signal that Cosmos has crossed from research artifact to foundational infrastructure.

3. The community is rigorously charting the next frontier

One of the hallmarks of a maturing field is the willingness to invest in rigorous evaluation, not to expose weaknesses, but to define the roadmap for what comes next. That’s exactly what several papers in this series do.

QUANTIPHY [3] (Stanford) shows that no current VLM, including the best proprietary and open-source models, can reliably perform quantitative physical estimation. SpatiaLQA [4] reveals that multi-step spatial reasoning remains a field-wide challenge when the logic gets complex. SRPO [22] (Fudan) explores using Cosmos-Predict2 for reward shaping in a zero-shot setting on domain-specific tasks and finds that bridging general video generation to task-specific policy optimization requires more than visual fidelity alone — a finding that applies equally to every foundation model in the space.

What makes these papers valuable isn’t what they criticize, it’s what they enable. Nobody invests in building a comprehensive benchmark for a model they don’t take seriously. The fact that teams at Stanford, Zhejiang, and Fudan are choosing Cosmos as the reference point for evaluating Physical AI confirms its position at the center of the field. And the research gaps these benchmarks expose, quantitative reasoning, compositional spatial logic, domain-specific adaptation, are precisely the areas where the next wave of innovation will happen, often building on top of the Cosmos platform itself (as PhyCo [17] and PhyCritic [2] already demonstrate).

Where Should You Turn Your Eyes in the Next Year?

After a week inside these papers, here’s where I think the highest-impact opportunities are. Not just for researchers, for engineers, product builders, and anyone trying to understand where this wave is going.

The sim-to-real gap is closing, but only for structured environments

SimScale [10], HorizonForge [11], and DriveLaW [12] show that simulation is becoming genuinely competitive with real-world data for autonomous driving. Policies trained on synthetic scenarios can match those trained on real miles. This is a game-changer, but it only works because roads are structured. Lanes, traffic lights, lane markings, predictable geometry.

The open frontier: unstructured environments. Agriculture, construction, mining, disaster response, off-road terrain. I surveyed 57 world model papers across CVPR 2026 and NeurIPS 2025. Zero apply world models to off-road environments. The research machinery is ready. Someone just needs to point it at the dirt.

Robots need to learn from failure, not just success

ViFailback [7] is one of the most underappreciated papers in this entire set. The idea that a robot should have a specialized failure diagnostic system, not just a policy that retries, is profound. Today’s VLA models are like students who only learn from correct answers. The next generation needs to learn from wrong ones too.

Watch for: failure-aware training, physics-grounded critic models like PhyCritic [2], and self-improving policies that get better from their mistakes. SRPO [22]’s self-referential approach is early, but the direction is right.

Efficiency matters as much as capability

DeltaWorld [16] shows you can match a 12B-parameter model with a fraction of the compute by rethinking representation. MAD [13] shows you can build competitive driving world models without industrial-scale budgets. AdapTok [18] adapts tokenization to scene complexity instead of wasting tokens on static frames.

The lesson: the next breakthroughs in Physical AI won’t just come from bigger models, they’ll come from smarter representations. Edge deployment, real-time inference, on-device world models for field equipment, this is where efficiency research meets industrial reality.

Data is the real bottleneck, and creative solutions are emerging

SceneScribe-1M [19] provides a million annotated videos. Lifting Unlabeled Data [21] turns internet video into 3D training data automatically. Real2Edit2Real [9] generates 50 demonstrations from 5. DAVIS-Ag produces 502K synthetic orchard images.

The pattern: the community is inventing new ways to generate training data at scale without expensive human annotation. But most of this data is still indoor or on-road. The datasets for off-road, agricultural, and industrial environments are just starting to appear (STONE, URScenes, TerraSeg). Building these datasets is one of the highest-impact things a researcher or company can do right now.

Multi-modal, multi-view, multi-task is the new normal

PlenopticDreamer [15] generates consistent video across multiple camera views. SceneScribe-1M [19] annotates video with text, depth, camera parameters, and 3D tracks simultaneously. DriveLaW [12] unifies prediction and planning in a single latent space. The era of single-task, single-view models is ending.

The implication: if you’re building physical AI systems, plan for multi-modal input from day one. LiDAR + camera + radar + IMU isn’t optional, it’s the minimum viable perception stack for anything that operates outside a controlled environment.

The Bigger Picture: 22 Papers, One Field, One Message

Here’s what I keep coming back to after writing this series:

CVPR is not easy to digest. Four thousand papers. Dozens of workshops. Hundreds of posters. The content is dense, technical, and scattered across subfields that don’t always talk to each other. That’s why I wrote these five posts, not to replace reading the papers, but to make them more readable for all of us. To connect the dots between a benchmark paper from Stanford and a robotics paper from Peking University and a driving simulation from Sorbonne.

The next wave of AI is physical. Not just language, not just images, not just video. AI that understands gravity, friction, deformation, and causality. AI that can act in the real world, not just describe it. The 22 papers in this series are the early evidence of that transition.

The unstructured world is the last frontier. We’ve made enormous progress on highways, indoor spaces, and urban environments. But the world doesn’t stop at the pavement. Agriculture, construction, mining, ports, disaster response, these domains represent hundreds of billions of dollars in economic activity and millions of vehicles that are rapidly adopting autonomy. The research community hasn’t caught up yet. That’s both a gap and an opportunity.

Where should you turn your eyes? Toward the physical. Toward the unstructured. Toward the systems that learn from failure, not just success. Toward efficiency, not just scale. Toward the infrastructure, the datasets, tokenizers, and benchmarks, that makes everything else possible.

The next year in Physical AI won’t be defined by who builds the biggest model. It will be defined by who builds the most useful one, the one that works in the mud, the dust, the rain, and the real world.

Let’s go build it.

The Full Series

If you are attending CVPR 2026 this week, follow my footsteps, here is part of my mandatory schedule:

• Part 1 — Can AI Understand Physics? Physical reasoning benchmarks
• Part 2 — Teaching Robots to Think, Plan, and Recover Embodied intelligence
• Part 3 — The Road Ahead Driving world models
• Part 4 — Seeing the Future Video generation and physics
• Part 5 — Building the Foundation Infrastructure and datasets

Full map of papers referencing NVIDIA Cosmos at CVPR 2026.

What is next?

The next CVPR, during CVPR 2027 we will see the impact on the community after the launch this week of Cosmos 3. Be ready to create the next wave in Physical AI using Cosmos as a foundation. What will be the impact you want to create by CVPR 2027?

I’m excited to share more about my journey in the intersection between Physical AI and multiple industries. If you’re interested in following along as I dive deeper into computer vision and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

References

Part 1 — Physical Reasoning Benchmarks

[1] F. Zhou, J. Huang, J. Li, D. Ramanan, and H. Shi, “PAI-Bench: A Comprehensive Benchmark For Physical AI,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[2] T. Xiong, S. Wang, G. Liu, Y. Dong, M. Li, H. Huang, et al., “PhyCritic: Multimodal Critic Models for Physical AI,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[3] L. Puyin, T. Xiang, E. Mao, S. Wei, X. Chen, et al., “QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[4] Y. Xie, X. Zhang, Y. Shan, H. Zhu, R. Tang, R. Wei, et al., “SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

Part 2 — Embodied Intelligence

[5] H. Tan, P. Co, Y. Xu, S. Rong, Y. Ji, C. Chi, X. Chen, et al., “Action-Sketcher: From Reasoning to Action via Visual Sketches for Robotic Manipulation,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[6] P. Xu, J. Zheng, and Y. Mu, “RoboAgent: Chaining Basic Capabilities for Embodied Task Planning,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[7] X. Zeng, X. Zhou, Y. Li, J. Shi, T. Li, L. Chen, L. Ren, et al., “Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols (ViFailback),” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[8] W. Huang, Y.-W. Chao, A. Mousavian, M.-Y. Liu, et al., “PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[9] Y. Zhao, H. Fan, D. Chen, S. Chen, L. Chen, X. Li, et al., “Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

Part 3 — Driving World Models

[10] H. Tian, T. Li, H. Liu, J. Yang, Y. Qiu, G. Li, J. Wang, et al., “SimScale: Learning to Drive via Real-World Simulation at Scale,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[11] Y. Wang, F. Pittaluga, Z. Tasneem, C. You, et al., “HorizonForge: Driving Scene Editing with Any Trajectories and Any Vehicles,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[12] T. Xia, Y. Li, L. Zhou, J. Yao, K. Xiong, H. Sun, et al., “DriveLaW: Unifying Planning and Video Generation in a Latent Driving World,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[13] A. Rahimi, V. Gerard, E. Zablocki, M. Cord, et al., “MAD: Motion Appearance Decoupling for Efficient Driving World Models,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

Part 4 — Video Generation & World Modeling

[14] Z. Ren, Y. Wei, X. Yu, G. Luo, Y. Zhao, B. Kang, et al., “VideoWorld 2: Learning Transferable Knowledge from Real-world Videos,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[15] X. Fu, S. Tang, M. Shi, X. Liu, J. Gu, M.-Y. Liu, D. Lin, et al., “Plenoptic Video Generation (PlenopticDreamer),” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[16] T. Kerssies, G. Berton, J. He, Q. Yu, W. Ma, et al., “A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens (DeltaWorld),” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[17] S. Narayanan, Z. Jiang, S. Narasimhan, et al., “PhyCo: Learning Controllable Physical Priors for Generative Motion,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

Part 5 — Infrastructure & Foundations

[18] Y. Li, C. Tian, R. Xia, N. Liao, W. Guo, J. Yan, H. Li, et al., “AdapTok: Learning Adaptive and Temporally Causal Video Tokenization in a 1D Latent Space,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[19] Y. Wang, K. Zheng, J. Wang, M. Chen, D. Novotny, et al., “SceneScribe-1M: A Large-Scale Video Dataset with Comprehensive Geometric and Semantic Annotations,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[20] Y. Haitman, A. Efraim, and J. M. Francos, “C-GenReg: Training-Free 3D Point Cloud Registration by Multi-View-Consistent Geometry-to-Image Generation,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[21] Y. Chen, Y. Zhang, H. Yu, J. He, Y. Wang, J. Huang, et al., “Lifting Unlabeled Internet-level Data for 3D Scene Understanding,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

[22] S. Fei, S. Wang, L. Ji, A. Li, S. Zhang, L. Liu, J. Hou, et al., “SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models,” in Proc. IEEE/CVF CVPR, Denver, CO, USA, 2026.

Author: Paula Ramos, PhD

Two months ago, I joined NVIDIA. And if I had to describe the experience so far in a few words, it would be: intense, inspiring, and deeply meaningful. But nothing quite prepared me for experiencing GTC from the inside.

We can describe GTC in multiple words, the most important tech conference in the world, even more important than CES or LEAP. The Super Bowl of AI. I can describe it as Mind-blowing epicenter of actual and future tech. GTC is a signal of where computing is heading, what problems we are prioritizing, and how entire industries and ecosystems are being reshaped in real time.

This year’s keynote, delivered by Jensen Huang, was a story about the evolution and reinvention of computing. With great vision about the future.

And as I reflect on both my first two months and what I saw at GTC, a few ideas stand out.

From CUDA to AI Factories

Innovation is a long Game. One of the most powerful moments in the keynote was the reflection on CUDA’s 20-year journey. Twenty years!!!

In an industry that often celebrates overnight success, and faces changes every week, this is a reminder that the most impactful platforms are built through long-term commitment. CUDA enabled GPUs for sure, and also created an ecosystem, a developer movement, and ultimately the foundation for modern AI.

What struck me most is how that same philosophy continues today. NVIDIA has been building chips, and platforms. And now we build AI factories.

The idea of AI factories reshape everything, data centers are production systems for intelligence, not just infrastructure. Tokens become the output. Throughput and efficiency become business metrics. And optimization is no longer optional, it is mandatory, not just for system performance also for energy consumption.

This shift is profound and shows how AI is becoming an industrial capability.

The Inference Era: When AI starts doing real work

Another key topic was the transition from training to inference. For years, the focus in AI was on training bigger and better models. But now, we are entering what Jensen described as the inference inflection point.

AI is no longer just learning. It is for reasoning, planning, generating, and increasingly… acting.

Agentic AI systems are emerging, systems that can break down problems, use tools, iterate, and execute tasks. We are moving from asking: “What is the answer?” To asking: “Can you do this for me?”

And that shift has massive implications for compute, software design, and productivity. Maybe our lives won’t be the same in a few months, when common people can use agents to assist in more single labors.

Structured + Generative: The future of trustworthy AI

One of the most important insights for me personally was the emphasis on combining structured data with generative AI. For a long time, enterprise systems relied on structured data, the “ground truth” of business. Meanwhile, generative AI unlocked the power of unstructured data.

Now, these two worlds are converging. And this is critical.Now, the future of AI is about generating reliable, contextual, and actionable intelligence. Across industries, healthcare, finance, manufacturing, this fusion will define how AI moves from experimentation to production.

A Remarkable Moment: OpenClaw, NemoClaw, and Open Source at Scale

One moment that truly stood out was the introduction of OpenClaw and NVIDIA’s NemoClaw.

At first glance, it might look like just another tool in the growing agent ecosystem. But not, OpenClaw represents something we don’t see often at this scale: open source becoming the foundation of the next computing paradigm, like Linux enabled modern computing, or Kubernetes enabled cloud-native systems. OpenClaw is emerging as a foundation for agentic systems.

And what made this moment great is that NVIDIA is joining to the Open source initiative to make this more secure, entreprise-ready and scalable with NVIDIA NemoClaw.

This is what great ecosystems look like: open innovation at the core, industry collaboration on top, and shared momentum across the community.

It’s a reminder that even in a world of trillion-dollar infrastructure and massive compute, we still build the future together.

Ecosystems over products

A clear GTC’s message is that none of this is happening in isolation. Everything is connected: Cloud providers, startups, enterprises, researchers, robotics companies, and open-source communities are all part of the same acceleration loop.

This creates a flywheel: 1) more developers → more innovation, 2) more innovation → more applications, 3) more applications → more adoption, 4) scaling technology → scaling collaboration.

Olaf on Stage: The moment that stayed with me:

And then… something unexpected happened. After a deeply technical keynote , discussions about AI factories, inference, agents, and open ecosystems, Jensen brought out Olaf!!! Yes, Olaf. The snowman from Frozen. I love it!

This was a symbolic moment that represents the transition from digital AI to physical AI. For years, AI lived in screens, and now it’s stepping into the real world with clear evidence

Olaf showed us a glimpse of a future where robots are trained in simulation, AI interacts with physical environments, and technology becomes something we can see, touch, and experience.

It also reminded me of something important: Even at the forefront of the most advanced technology, we are still building for humans. For creativity.
For storytelling. For connection. And show me how we can translate tokens into a real future.

A Personal Reflection

Two months into NVIDIA, I feel incredibly grateful. I am really grateful to be learning, be challenged, and be part of a company helping shape what comes next.

This moment is not just for a few companies or a specific group of developers. This is a platform shift. And with every platform shift comes opportunity: to build, learn, and rethink what’s possible.

Because the future is not something we wait for, it’s something we build.

What is next?

I’m excited to share more about my journey in the intersection between Physical AI and multiple industries. If you’re interested in following along as I dive deeper into computer vision and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

When I began my journey as a Latina in tech, I didn’t just enter the world of AI, I stepped into an opportunity to expand it.

I learned the math, the models, and the systems. But I also learned that visibility is powerful. Every time one of us speaks, publishes, mentors, or leads, we make the space larger for the next woman walking in.

That realization didn’t intimidate me. It energized me. It shaped how I chose to grow, and how I chose to give back.

Over the years, that commitment has taken many forms: organizing workshops at CVPR, mentoring early-career researchers, supporting first-time authors, and building technical solutions that bring AI into rigorous, real-world sectors. I’ve worked on making research not just innovative, but applicable, bridging anomaly detection, foundation models, and computer vision systems into domains like agriculture and other production environments where robustness, reliability, and impact truly matter.

Because for me, advancing AI has never been just about publishing result it has been about ensuring that research translates into meaningful, deployable solutions.

More than anything, this journey has strengthened my belief in one thing:

Community changes trajectories. And that’s exactly why I am deeply committed to the Women in Computer Vision (WiCV) Workshop.

In this post, I want to share what I’ve learned — not just about writing stronger papers, reviewing thoughtfully, or mentoring effectively — but about how each of us can actively shape the field by choosing to participate.Learning the Field and the Gaps

As my career in AI and computer vision grew, so did my exposure to the breadth of the field. I worked across research, applied AI, and industry-driven problems, organizing workshops at CVPR on topics like visual anomaly detection, foundation models, and AI applications in agriculture. but again it was beyond the Women in the Street series of videos, or helping researcher with workshops, tutorials or papers. It all was about to bridge the gap between knowledge and adopters.

Agriculture, in particular, became a powerful learning space for me. It is complex, underexplored, deeply tied to real-world impact, and full of challenges that computer vision can help solve, if we approach them thoughtfully. Working in this space taught me how important it is to bridge theory and practice, and how much stronger solutions become when diverse perspectives are involved.

Yet, no matter the topic, the pattern repeated itself: I was often still the only woman in the room.

Turning Experience Into Intention

At some point, I realized that simply succeeding individually was not enough. If access to information, mentorship, and visibility had been hard for me to find, then it was likely even harder for others coming after me.

That’s when my focus shifted: I started mentoring more intentionally. I supported first-time authors. I encouraged women to submit papers even when they doubted themselves. I volunteered time to review, organize, and create spaces where learning felt possible, not intimidating.

Being part of the Women in Computer Vision (WiCV) community at CVPR is a natural extension of that intention.

This workshop is not just about publishing papers. It’s about lowering barriers, sharing knowledge, and building confidence, together.

An Open Invitation

If you are reading this and wondering whether there is space for you, let me be clear… There is, whether you are:

• Submitting your first paper or abstract — Call for Papers
• Reviewing for the first time — Call for Reviewers
• Mentoring students or early-career researchers — Call For Mentors
• Supporting the community as a sponsor — Send an email to wicvcvpr2026@googlegroups.com

Your participation matters. Communities don’t grow because one person speaks louder, they grow because many voices choose to show up.

How to Submit Better Papers — Call for Papers

If you’re preparing a submission, here are some practical guidelines that can make a real difference:

• Start with a Strong, Clear Abstract
  Your abstract is an invitation. Clearly state the problem, why it matters, and what your contribution is. Aim for clarity and excitement, not complexity.
• Connect to the Broader Field
  Show how your work fits into the current computer vision landscape. Cite relevant research and explain how your approach builds on or differs from existing work.
• Be Rigorous and Transparent
  Make your methodology reproducible. Explain your design choices and be honest about limitations. Strong papers are not perfect — they are clear and thoughtful.
• Get Feedback Early
  Share drafts with peers or mentors. External feedback often reveals gaps or strengths you can’t see yourself.

And most importantly: don’t self-reject. Workshops are meant for discussion, growth, and early ideas.

How to Be a Great Reviewer — Call for Reviewers

Reviewers shape the tone of a community. A good review doesn’t just judge, it guides.

• Be Constructive, Not Dismissive: Behind every submission is effort. Offer suggestions that help authors improve.
• Focus on Clarity and Evidence: Highlight where explanations are unclear or claims need support.
• Stay Fair and Objective: Evaluate the work on its merits, not on personal preferences or familiarity with the topic.

Thoughtful reviews make better papers, and a better field.

How to Be a Mentor — Call For Mentors

Mentorship doesn’t require perfection. It requires presence.

• Share Your Real Journey: Talk openly about challenges, not just successes. Authentic stories resonate.
• Listen Actively: Sometimes guidance starts with simply being heard.
• Encourage Action: Push mentees to take the next step: submit the paper, apply for the workshop, speak up.

Confidence grows through action, not waiting.

The Power of Community

The room is changing, but only because people choose to change it.

Every paper submitted, every review written with care, every mentoring conversation contributes to something larger than any individual achievement. Community transforms isolation into momentum, doubt into confidence, and ideas into impact.

You don’t need permission to belong here.

Let’s keep building — together. 💜

What is next?

I’m excited to share more about my journey in the intersection between Physical AI and multiple industries. If you’re interested in following along as I dive deeper into computer vision and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

A new chapter and the backpack we choose to carry

Before starting, choosing what to carry

Every new chapter feels like starting a journey. Before taking the first step, there is a pause, a moment to look back and ask: what am I carrying with me? A backpack is not just a practical object; it is a choice. Not everything fits, and not everything has the same weight. What we decide to pack shapes how we move forward.

In this moment of transition, I’ve focused less on titles or achievements, and more on the experiences, conversations, and lessons that have slowly filled that backpack over time.

When AI meets the real world

Much of my professional path has lived where artificial intelligence stops being a promising idea and starts working in the real world, where data is imperfect, environments change, and systems must adapt to stay useful.

There, I learned that good models alone are not enough. We also need strong workflows, an honest view of limits, and close collaboration between researchers, builders, and the people who use these systems every day. This, for me, is the heart of Physical AI.

What I pack for this new role

I start this new chapter at NVIDIA with excitement and humility. I see it not as a destination, but as a continuation of the path I’ve been walking. In my backpack, I carry the experience of working with diverse communities, seeing projects grow in real environments, and learning that real impact is built step by step.

My motivation remains the same: helping companies, users, and practitioners cross that last mile, turning powerful platforms into complete Physical AI systems that are reliable, scalable, and able to work even when conditions are not ideal.

The backpack fills itself when we give

Nothing in our backpack appears overnight. It fills with every shared conversation, every workshop, every small prototype, and every moment of listening and learning. Often quietly, but always with impact.

There’s no need to rush toward the result. When we give with intention, build with honesty, and keep learning, the backpack fills itself, and what we offer to the world often finds its way back.

Continuing the walk, with gratitude

This chapter is not about titles or transitions. It’s about responsibility, to the people building these systems and to the real world where they operate.

I start this journey with my backpack ready and open to change, grateful to learn, contribute, and walk alongside a community that believes the future of AI is not only intelligent, but truly tangible.

Moving forward, with gratitude 🚀

Starting a new journey with a backpack full of learning, purpose, and gratitude.

What is next?

I’m excited to share more about my journey in the intersection between Physical AI and multiple industries. If you’re interested in following along as I dive deeper into computer vision and continue to grow professionally, feel free to connect or follow me on LinkedIn. Let’s inspire each other to embrace change and reach new heights!

Visualizing What’s Next

“Visualize what is next. The cycle you’ve been following is reaching its fulfillment. You’re at the finish line of your goals, and it’s time to level up.”

As I sit down to write this final blog post, wearing my Voxel51 hat (literally the same hat I wore on my first day), I’m filled with a profound sense of gratitude and satisfaction. This moment feels like the reflective culmination of a deeply meaningful journey, one that has transformed me both professionally and personally.

From New Beginnings to Fulfillment

At the start of any new chapter, there’s a natural mix of excitement and nerves. When I joined Voxel51, I wrote my first blog post about stepping into the unknown. Looking back now, I can clearly see how that initial curiosity, paired with strong community support, evolved into a journey defined by growth, purpose, and impact.

More Than a Workplace

Voxel51 has been more than just a workplace; it has been a community of incredibly talented people passionate about pushing the boundaries of computer vision. I want to give a special shout-out to the amazing Community Team I had the privilege of working with.

The way our DevRel team collaborates across partners, developers, customers, and internal teams is genuinely inspiring. Their work adds tremendous value, not only to the product but to the broader ecosystem, and made my time at Voxel51 especially rewarding.

Making an Impact with FiftyOne

For those who may not know, Voxel51’s product, FiftyOne, is a powerful platform that goes far beyond data visualization. It unifies multimodal datasets, helps teams deeply understand and curate their data, and enables model retraining for better performance. It directly addresses real and persistent challenges in the computer vision community.

Over the past year, I’ve had the opportunity to work on projects that were warmly received both inside and outside the company. We focused on real-world industry pain points across manufacturing, healthcare, agriculture, and video, showcasing how FiftyOne fits into end-to-end workflows.

This work wasn’t just about fixing annotations or cleaning datasets. It was about building a vibrant ecosystem where researchers can create plugins, share innovations, and where teams across marketing, sales, product, and engineering truly understand and communicate the value of AI and computer vision.

Turning Vision into Impact

As I reflect on this chapter, I’m proud of the tangible impact we created together. What started as ideas and conversations turned into real momentum across the community, industry, and global engagement:

• 30+ talks delivered across meetups, conferences, and community events
• 6+ hands-on workshops, focused on practical, end-to-end workflows
• 4+ keynote speaker, representing Voxel51 in broader industry/academic conversations
• 1250× increase in awareness for industry-related content showcasing the value of FiftyOne. Adding industries such as manufacturing, healthcare, agriculture, and video
• 15+ repositories and one-click, end-to-end demos, designed to lower the barrier to adoption
• Multiple getting-started experiences to help users reach their first success faster
• Creating blogs to help tech and non-tech users understand and be on top of the latest AI trends and academic results

Beyond technical content and code, community building was at the heart of this journey:

• Built En Español events, creating inclusive, global, and welcoming spaces for diverse voices
• Launched a Women in AI community, with 200+ attendees on average per session
• Reaching 100+ speakers across multiple countries and AI topics
• Created the Best of Conferences series, opening global discussions with researchers
• Opened and supported communities in new places, including Raleigh, Iowa, San Diego, Valencia, Madrid
• Organized international tours, including the Spain tour and a Germany tour, moments where the energy of the community made me feel, at times, like a rock star in the best possible way
• Attended 5+ conferences, always seeking new speakers, collaborators, and voices. Organizing tutorials, demos, and workshops at academic conferences.

And behind every visible outcome was the invisible work: finding venues, ordering food, coordinating logistics, supporting speakers, and, at the same time, creating technical content that delivered real value to every attendee.

I’m not sure how many miles I traveled along the way, but I know it’s enough to say this chapter was built with intention, energy, and care.

A Community That Truly Engages

What I’m most proud of is the community spirit at Voxel51. From internal retreats to open-source contributions, this company has built a strong remote culture while keeping people genuinely connected. Moments with our Voxel51 community reinforced for me how intentional and impactful this community truly is.

Looking to the Future

To the entire computer vision community: pay attention to Voxel51. If you haven’t yet explored FiftyOne, their open-source platform, I highly recommend you do. The company is agile, constantly innovating, and continuously bringing together the brightest minds in the field through its events. In 2026 and beyond, Voxel51 will continue to face new challenges, and the best way to stay informed is to be part of the community: participate in events, engage with the open-source ecosystem, and learn alongside truly brilliant people.

As I close this chapter, I want to thank everyone who made this journey so rewarding. It has been a privilege to grow alongside such a passionate team and to be part of a community that firmly believes in the power of shared vision.

Thank you, Voxel51, for the trust, the collaboration, and the experience. As I embark on my next path, I take with me the lessons learned, the friendships forged, and the inspiration we have built together.

Here’s to new horizons and to the continued journey of growth and discovery for all of us.

What is next?

Stay tuned! https://www.linkedin.com/in/paula-ramos-phd/