Embodied foundation models need loop data.

World-action model framing where future visual states and actions are predicted in an aligned policy model.

NVIDIA's current open humanoid VLA release within an end-to-end Isaac workflow for teleoperation, training, evaluation, and deployment.

A hierarchical whole-body VLA for continuous humanoid locomotion, manipulation, balance, tactile feedback, and long-horizon autonomy.

A masked depth model for refining incomplete sensor depth into metric geometry for reconstruction, tracking, and manipulation.

A feed-forward 3D foundation model for streaming scene reconstruction, camera-pose estimation, and long-sequence point-cloud generation.

An autoregressive video-action world-model policy that interleaves future video-latent prediction with robot action generation.

A family of vision foundation encoders pretrained for dense spatial perception and downstream embodied understanding.

A 4B pragmatic VLA foundation model pretrained on roughly 20,000 hours of real-world data from nine dual-arm robot configurations.

A whole-body VLA release focused on broader cross-embodiment transfer, mobile manipulation, and predictive dynamics supervision.

An open interactive world simulator derived from video generation, with camera- and action-conditioned variants and long-horizon generation.

World-model architecture using looped transformer refinement over latent states for iterative environment understanding.

Object-addressable world-action model that decomposes scenes into robot and object slots while jointly predicting future world state and actions.

A modular, MIT-licensed VLA development stack with multiple VLM backbones, model scales, and dataset integrations.

An open 4B VLA pretrained across more than 20 embodiments, with physical-hardware evaluation of pretrained robotic capability.

A unified embodied model using interleaved vision-text-action pretraining for perception, reasoning, planning, and continuous robot control.

A dual-system VLM plus VLA model for planning and fine-grained control on long-horizon mobile-manipulation tasks.

Embodied reasoning and action model family that brings Gemini capabilities into robotics through vision, language, and physical action.

An agentic robotics model combining advanced multimodal reasoning with vision-language-action control for multi-step physical tasks.

Open foundation model for generalist humanoid robots, focused on whole-body and manipulation behavior from multimodal robot data.

World foundation model platform for physical AI, built around predictive world modeling and data processing workflows.

An optimized OpenVLA fine-tuning recipe with continuous action chunks, multi-image input, and faster high-frequency control.

A VLA model designed for open-world generalization by combining heterogeneous robot data, semantic subtask prediction, and high-level knowledge transfer.

Compact open vision-language-action policy designed for practical robot fine-tuning and deployment through the LeRobot ecosystem.

A spatially enhanced 4B VLA pretrained on 1.1 million real-robot episodes with explicit geometry-aware representations.

A 0.9B soft-prompted VLA that adapts a shared backbone across heterogeneous embodiments and action spaces.

Large-scale foundation world model for generating action-controllable interactive environments from visual prompts.

Open-source generalist robot policy pretrained on Open X-Embodiment trajectories and designed for fine-tuning to new robots and tasks.

Open-source vision-language-action model for generalist robotic manipulation, trained on diverse real-world robot demonstrations.

Generalist robot policy family and open-source robotics model package from Physical Intelligence.

Diffusion foundation model for bimanual manipulation that uses large-scale robot data to generate action trajectories.

Action Chunking with Transformers predicts short action sequences for efficient imitation learning in manipulation tasks.

Visuomotor policy approach that represents robot behavior as a conditional denoising diffusion process.

Self-improving generalist robotic agent that collects new demonstrations to improve its own manipulation capabilities.

Vision-language-action model that transfers web-scale vision-language knowledge into robotic control.

Cross-embodiment RT model family trained from the Open X-Embodiment mixture to study transfer across robots and tasks.

Interactive real-world simulator research that models how visual scenes change under actions and interaction.

Robotics Transformer policy trained on large-scale real-world robot demonstrations for language-conditioned manipulation.

Language-grounded robotics approach that combines language-model planning with affordance scores from robot skills.