A Practical Map of the Models That Turn Perception Into Robot Action

A structured map of Physical AI models covering VLA policies, embodied reasoning, world models, manipulation, locomotion, tactile learning and simulation.

Introduction

A language model can describe how to load a dishwasher without producing a single motor command. A robot policy can output joint actions without understanding why the plate belongs in a rack. A world model can predict what a future scene may look like without controlling real hardware. Physical AI is the stack that connects these separate functions: sensing, representation, reasoning, planning, action generation, low-level control, safety and recovery.

The term is often used so broadly that simulators, video generators, perception networks and complete robot brains appear in one undifferentiated list. This map separates them by input, output and role. It covers widely used or strategically important systems available by July 11, 2026, including NVIDIA Isaac GR00T 1.7, Gemini Robotics, Gemini Robotics-ER 1.6, π0 and π0.5, OpenVLA, Octo, RT-2, Figure Helix 02 and 1X Redwood. “Open” is also divided into weights, code, data and license because those are not the same form of access.

Key findings

  • A VLA model converts visual observations and language instructions into actions; an embodied reasoning model may plan and verify without directly controlling joints.
  • World models generate or predict future states and data. They support planning and training but are not automatically deployable robot policies.
  • OpenVLA, Octo, openpi and NVIDIA’s GR00T releases provide different combinations of code, weights and licenses; none is plug-and-play across every robot.
  • Figure Helix and 1X Redwood are vertically integrated proprietary policies tied to each company’s hardware and data pipeline.
  • The largest missing layer in public benchmarks is long-duration recovery: most evaluations measure task success, not intervention rate across real shifts.

Major Physical AI models and their roles

Architecture details are limited to what the organization publishes. “Closed” means deployable weights are not publicly released.

ModelOrganization / dateCategoryInputs → outputsHardware or task evidenceAccess and main limitation
Isaac GR00T 1.7NVIDIA; July 2026Open reasoning VLA / humanoid policyImages, language and robot state → action trajectoriesHumanoid manipulation, finger control and cross-embodiment workflows through Isaac Lab and ROSWeights and code available under NVIDIA terms; still requires robot-specific data, tuning and safety integration
Gemini Robotics 1.5Google DeepMind; current 2026 pageVLA / multimodal controlImage, text, audio or video context → robot actionsDexterous manipulation and multi-step physical tasks on several robot formsClosed service and selected access; hardware latency, cost and deployment details vary
Gemini Robotics-ER 1.6Google DeepMind; 2026Embodied reasoning and orchestrationMulti-view perception, language and tools → plans, spatial references and success decisionsTask planning, instrument reading, pointing and retry/progress logicAvailable through Google tooling; it is a high-level brain, not a universal low-level controller
Gemini Robotics On-DeviceGoogle DeepMind; 2026On-device VLAImage, text and action context → actionsLocal dexterous manipulation and adaptationPrivate preview / trusted testers; model weights and broad benchmarks are not public
π0Physical Intelligence; October 2024Generalist VLA policyImages, language and proprioception → continuous actionsMulti-robot manipulation including laundry, boxes and table tasksBase weights and code through openpi; substantial fine-tuning and hardware integration required
π0.5Physical Intelligence; April 2025Open-world VLA policyVisual observations, language and robot state → actionsGeneralization to new homes and semantically varied choresResearch release details are more limited than openpi; broad independent reproduction remains sparse
OpenVLA 7BStanford, UC Berkeley and collaborators; June 2024Open-source VLAImage and language → tokenized robot actionsBridgeData V2 and LIBERO manipulation evaluationsCode and weights open; action representation and embodiment transfer need adaptation
OctoUC Berkeley and collaborators; 2024Generalist robot policyImages, language and task observations → action chunksFine-tuning across multiple robot datasets and embodimentsOpen code and checkpoints; performance depends heavily on dataset and embodiment match
RT-2Google DeepMind; July 2023VLAImages and text → action tokensSemantic manipulation and web-knowledge transfer on robot tasksClosed model; useful historical foundation but not a downloadable general platform
Helix 02Figure AI; 2026Proprietary full-body VLA / control hierarchyRobot cameras, language and state → upper-body and locomotion actionsFigure 03 dishwasher, room-tidying and industrial tasksClosed and tied to Figure hardware; benchmarks and intervention logs are company-controlled
Redwood AI1X Technologies; 2025-2026Proprietary home-robot policyVision, language and robot state → NEO actionsHousehold skills and data collection on NEOClosed; Expert Mode means human assistance remains part of product operation
Skild BrainSkild AI; current 2026Cross-embodiment robotics foundation modelMultimodal observations and instructions → robot policiesClaims across manipulation, locomotion and different embodimentsClosed commercial access; public architecture and benchmark detail remain limited
Cosmos world foundation modelsNVIDIA; 2025-2026World models and synthetic dataVideo, text and physical scene context → predicted or generated video/world statesData generation, scenario expansion and robot training supportModels and tools vary by release; output is not a motor policy without a downstream controller
Open X-Embodiment / RT-XGoogle DeepMind and 20+ institutions; 2023Dataset mixture and cross-robot policy researchMulti-robot trajectories → training data for generalist policiesShared data from many robot embodiments and RT-X experimentsDataset heterogeneity, licenses and action normalization complicate reuse

Taxonomy from data to motion

This text structure can be converted directly into an infographic.

LayerQuestion answeredExamplesOutput passed downstream
Data and simulationWhat experiences can the system learn from?Open X-Embodiment, Isaac Lab, teleoperation datasets, Cosmos-generated scenariosTrajectories, labels, synthetic scenes and demonstrations
Perception and tactile representationWhat objects, contacts and body states exist now?Vision encoders, depth models, Sparsh-style tactile representationsFeatures, poses, contact state and uncertainty
World modelWhat may happen next?NVIDIA Cosmos and learned dynamics modelsPredicted future states or generated rollouts
Embodied reasoningWhat is the task, constraint and next subgoal?Gemini Robotics-ER and high-level plannersPlan, spatial grounding, tool call or success judgment
VLA / action policyWhich action sequence should execute the instruction?GR00T, Gemini Robotics, π0, OpenVLA, Octo, Helix, RedwoodEnd-effector targets, action chunks or joint-level commands
Low-level manipulation and locomotionHow should motors track the target while maintaining contact and balance?Whole-body controllers, reinforcement-learning locomotion policies, impedance controlTorque, position or velocity commands
Safety, recovery and supervisionShould the robot continue, retry, stop or ask for help?Collision limits, fall detection, remote assistance and success detectorsStop, retry, reduced-speed mode or human handoff

Vision-language-action models are policies, not complete robots

A VLA model receives visual observations and an instruction, then predicts actions. Those actions can be Cartesian gripper targets, joint commands or tokenized trajectories. The model still depends on calibrated cameras, robot kinematics, a real-time controller, collision handling and a safe execution layer. A checkpoint trained on one arm cannot be assumed to control a biped because action spaces, timing and body dynamics differ.

OpenVLA, Octo and openpi give researchers inspectable starting points. NVIDIA GR00T adds a broader humanoid development workflow around teleoperation, simulation, training, evaluation and deployment. Gemini Robotics and proprietary systems such as Helix integrate reasoning and action more tightly but expose less implementation detail.

Embodied reasoning sits above continuous control

Gemini Robotics-ER 1.6 is a useful example of a high-level model. It reasons about multi-view scenes, selects subgoals, interprets instruments and checks whether a step succeeded. It can orchestrate tools and decide to retry. That output can guide a policy, but it is not a replacement for a stable controller running at motor frequency.

This distinction matters when companies call a model a robot brain. A planner operating once per second and a torque controller operating hundreds or thousands of times per second are both necessary, yet they solve different problems. Public model maps should label the temporal and control level instead of placing every model in one “intelligence” box.

World models produce futures, not necessarily actions

World models learn how scenes may evolve. They can generate synthetic training clips, predict the effect of a candidate action or expose rare failure states in simulation. NVIDIA Cosmos is positioned as a family of world foundation models supporting physical AI data and simulation. A generated future can improve a policy, but the model does not become a deployable controller unless a planner or action policy consumes that prediction.

Video-generation quality is also not physical validity. Contact forces, friction, object mass and actuator limits can be visually plausible while mechanically wrong. Real-robot validation remains necessary, especially for dexterous manipulation and dynamic balance.

Manipulation, locomotion and touch need specialized layers

Generalist VLAs focus attention on semantic tasks, yet hand contact and biped balance are sensitive to milliseconds, force and local geometry. Many systems therefore pair a large model with specialized manipulation controllers, tactile estimators and reinforcement-learning locomotion policies. The large model selects an objective; lower layers execute it under physical constraints.

Tactile learning remains underrepresented in public foundation-model maps. Camera-based touch sensors can produce dense images, while force-torque sensors provide lower-dimensional contact signals. The model must synchronize those streams with vision and proprioception. Few public benchmarks compare tactile generalization across different hands.

Open source has four separate meanings

A project can release a paper without code, code without weights, weights without training data or all three under a license that restricts commercial use. OpenVLA and openpi provide meaningful code and checkpoint access. GR00T publishes code and weights under NVIDIA terms. Gemini Robotics, Helix, Redwood and Skild Brain remain controlled commercial systems.

Dataset access is equally important. Open X-Embodiment demonstrated the value of pooled robot trajectories, but heterogeneous sensors, coordinate frames and licenses make mixture training difficult. A model being downloadable does not remove the cost of collecting task-specific demonstrations on the target robot.

Benchmarks still miss operational reliability

LIBERO, BridgeData evaluations and laboratory task suites measure progress, but factories and homes need longer metrics: tasks per hour, intervention rate, recovery time, damage, safety stops and performance after lighting or object changes. A model can score well on short episodes and still fail repeatedly during a shift.

The next useful comparison should report policy success together with robot embodiment, control frequency, scene distribution, number of trials and human recovery. Without those fields, claims of generality remain difficult to compare across organizations.

Limitations and missing information

  • The map prioritizes influential and currently accessible systems; it is not a list of every robotics model published worldwide.
  • Organizations disclose architecture, training data and benchmarks at different levels of detail.
  • Model names can cover several checkpoints or product versions, and access terms may change.
  • Simulation tools such as Isaac Lab are included only as pipeline layers, not mislabeled as robot policies.
  • No common benchmark compares all models on the same hardware, tasks, intervention rules and safety constraints.

Conclusion

Physical AI is a pipeline, not one giant model. Perception encodes the scene. A world model estimates possible futures. Embodied reasoning turns a goal into subgoals and checks progress. A VLA or specialized policy produces actions. Low-level controllers maintain contact, balance and timing. Safety and supervision decide when to stop or ask for help. Systems such as GR00T, Gemini Robotics, π0, OpenVLA, Helix and Redwood occupy different parts of that stack even when their marketing language overlaps.

For developers, model choice begins with the output required by the robot. A high-level planner is useful when task decomposition is the bottleneck. An open VLA is useful when the team can collect embodiment-specific data and tune the action space. A world model is useful for prediction and synthetic experience. None removes the need for controllers, calibration, safety and field evaluation. The most informative 2026 model map therefore labels role, access, embodiment and evidence rather than declaring a single “best robot brain.”

Frequently asked questions

What is a Physical AI model?

A Physical AI model processes information about the physical world and contributes to perception, reasoning, planning, prediction or control. The term can include VLAs, world models, tactile encoders and locomotion policies. It should not imply that one model performs every layer from camera input to safe motor torque.

What is a vision-language-action model?

A VLA takes visual observations and language instructions and predicts robot actions. Actions may be tokenized, continuous end-effector targets or joint trajectories. A VLA still needs robot-specific calibration, control, safety and often fine-tuning. It is closer to a task policy than to a complete autonomous robot product.

Is a robotics world model the same as a robot policy?

No. A world model predicts or generates possible future states. A policy selects actions. A planning system can use a world model to compare candidate actions, but generated video or predicted state alone does not move a robot. Some systems combine both functions, so architecture and outputs must be checked.

Which Physical AI models are open source?

OpenVLA, Octo, openpi and NVIDIA Isaac GR00T provide substantial public code or weights, though licenses and completeness differ. Open X-Embodiment provides datasets and research tooling. Gemini Robotics, Figure Helix, 1X Redwood and Skild Brain are controlled-access or proprietary systems. Open code does not guarantee open training data.

Can the same model control any robot?

Cross-embodiment models aim to transfer knowledge, but robots differ in cameras, joints, hands, timing, payload and safety limits. Most models require action-space adapters, demonstrations and fine-tuning for a target platform. General semantic knowledge transfers more easily than precise contact-rich control or dynamic locomotion.

Why are simulators not classified as models?

A simulator computes how a virtual environment changes under actions. It can generate data and evaluate policies, but it is not itself the learned policy. Learned world models can overlap with simulation, yet tools such as Isaac Lab remain environments and workflows. Keeping the categories separate prevents software infrastructure from being counted as robot intelligence.

Sources and methodology

Models were classified by published inputs, outputs and role rather than by company labels. Primary model pages, repositories, model cards and official technical blogs were reviewed. Access was separated into code, weights, data and commercial availability.

Verified July 11, 2026. Version numbers are stated where the organization publishes them. Missing architecture, license or benchmark detail is marked closed or not publicly disclosed instead of inferred.

  1. Develop Humanoid Robot Policies End-to-End with NVIDIA Isaac GR00T — NVIDIA · July 2026
  2. Isaac GR00T platform — NVIDIA · accessed July 11, 2026
  3. Gemini Robotics — Google DeepMind · accessed July 11, 2026
  4. Gemini Robotics-ER 1.6 — Google DeepMind · accessed July 11, 2026
  5. Gemini Robotics On-Device — Google DeepMind · accessed July 11, 2026
  6. π0: Our First Generalist Policy — Physical Intelligence · October 31, 2024
  7. Open Sourcing π0 — Physical Intelligence · February 4, 2025
  8. π0.5: A VLA with Open-World Generalization — Physical Intelligence · April 22, 2025
  9. OpenVLA repository — OpenVLA project · accessed July 11, 2026
  10. Octo project — Octo project · accessed July 11, 2026
  11. RT-2 — Google DeepMind · July 28, 2023
  12. Helix 02 — Figure AI · 2026
  13. Redwood AI — 1X Technologies · accessed July 11, 2026
  14. Skild Brain — Skild AI · accessed July 11, 2026
  15. Open X-Embodiment repository — Google DeepMind and collaborators · accessed July 11, 2026

Related TechniaHQ guides

Official image recommendations

Fact-check report

Verified: July 11, 2026

Confirmed

  • GR00T 1.7, Gemini Robotics-ER 1.6 and current access states are tied to official pages.
  • World models, simulators and action policies are separated by output role.
  • Open access is split into code, weights, data and license.

Not confirmed or incomplete

  • Training dataset composition is incomplete for several closed models.
  • No common real-robot benchmark covers every model in the catalog.
  • Independent reproduction of proprietary home and humanoid policies is unavailable.

Fast-changing information

  • Model versions, licenses, preview access and public checkpoints can change within weeks.