THE MODELS DEFINING PHYSICAL AI

By Dr. Barak Or, Founder & CEO, STATE16

From vision-language-action systems to world models, these are the technologies giving machines the ability to perceive, reason and act in the physical world.

Physical AI represents the transition from artificial intelligence that works primarily with digital information to AI that can operate in the real world. A physical AI system must interpret camera images, understand language, process sensor data, reason about space, predict the consequences of its actions and control a physical machine.

There is currently no single model that can be described as the "GPT of robotics." Instead, the Physical AI ecosystem is developing around three main categories: Vision-Language-Action models, robot control policies and world models.

There is also no reliable global ranking of model adoption. The names below represent the most recognized and influential models across robotics research, open-source development and commercial humanoid robotics.

Vision-Language-Action Models

Vision-Language-Action models, commonly known as VLAs, connect perception and language directly to robot control. They receive visual information, natural-language instructions and robot-state data, then generate the actions required to complete a task.

NVIDIA Isaac GR00T 1.7 is one of the most important open models for humanoid robotics. It accepts images, language and robot-state information, and generates actions across different robot embodiments. NVIDIA released it under the Apache 2.0 license and integrated it into development workflows that support post-training, simulation and deployment on real robots.

Physical Intelligence's π0 family is another major force in general-purpose robot learning. The original π0 demonstrated a generalist policy trained across multiple tasks and robots. Later versions expanded its ability to operate in unfamiliar environments. The latest generation, π0.7, is designed to follow new language instructions, combine previously learned skills and perform tasks that were not explicitly included in its training data.

Google DeepMind's Gemini Robotics combines Gemini's multimodal intelligence with physical action. Gemini Robotics 1.5 translates images and instructions into motor commands, while Gemini Robotics-ER 1.6 focuses on embodied reasoning, spatial understanding, planning, task-completion detection and decision-making. Together, they represent a layered architecture in which one model reasons about the task and another controls the robot.

Figure's Helix 02 is a proprietary VLA system developed specifically for Figure's humanoid robots. It extends the original Helix architecture from upper-body manipulation to full-body autonomy, including walking, balancing and object manipulation. Figure describes it as a single neural system that controls the robot directly from visual input.

Open Models for Researchers and Developers

OpenVLA is one of the best-known open-source VLA models in academic robotics. It has 7 billion parameters and was pretrained on approximately 970,000 robot episodes from the Open X-Embodiment dataset. It can be adapted to new robots and tasks through parameter-efficient fine-tuning.

SmolVLA, developed within the Hugging Face LeRobot ecosystem, takes a different approach. At approximately 450 million parameters, it is designed to run on consumer hardware and can be trained on a single consumer GPU. This makes it particularly relevant for education, affordable robot arms and smaller robotics laboratories.

Octo is an open generalist robot policy based on a Transformer and diffusion architecture. It was pretrained on 800,000 robot episodes and released in 27-million and 93-million parameter versions. Its modular design allows developers to adapt it to new cameras, sensors, action spaces and robot platforms.

The Models That Shaped the Field

Google's RT-1, RT-2 and RT-X helped establish the modern VLA category. RT-2 demonstrated that robot actions could be represented as tokens inside a vision-language model. RT-X expanded robot learning across multiple institutions and embodiments using the Open X-Embodiment dataset, which contains more than one million real robot trajectories from 22 robot embodiments.

An important distinction is that Open X-Embodiment is a dataset and research initiative, not a single model. RT-1-X and RT-2-X are models trained using its multi-robot data.

Two additional approaches remain highly influential in practical robot learning. ACT, or Action Chunking with Transformers, predicts a sequence of future actions instead of producing one action at a time. Diffusion Policy generates smooth action sequences through a conditional denoising process and is especially effective when several different actions could successfully complete the same task.

World Models: Teaching AI How the Physical World Works

World models do not simply decide which motor command to execute. They learn how environments change over time, how objects interact and what may happen after a particular action.

NVIDIA Cosmos 3 combines physical reasoning, world generation and action generation in an open foundation model. It can support synthetic-data creation, scenario simulation, action prediction and the development of robots, autonomous vehicles and other physical AI systems.

Meta V-JEPA 2 learns predictive representations from video rather than generating every future pixel. After additional training on robot data, it can support planning for reaching, grasping and pick-and-place tasks in unfamiliar environments.

Google DeepMind's Genie 3 generates interactive environments from text descriptions. These environments can be explored in real time, making Genie relevant for simulation, embodied-agent research and the creation of training environments. It is primarily a world-generation system rather than a direct robot-control policy.

The Physical AI Model Map

• General-purpose VLAs: GR00T 1.7, π0.7 and Gemini Robotics. Primary role: turning perception and language into action.

• Humanoid control: GR00T 1.7 and Helix 02. Primary role: manipulation and whole-body autonomy.

• Open research models: OpenVLA, SmolVLA, Octo and RT-X. Primary role: experimentation, fine-tuning and deployment.

• Task-specific policies: ACT and Diffusion Policy. Primary role: learning reliable actions from demonstrations.

• World models: Cosmos 3, V-JEPA 2 and Genie 3. Primary role: prediction, simulation and synthetic data.

The Bottom Line

The future of Physical AI will probably not depend on one universal model. A complete robotic system is more likely to use several specialized layers: a reasoning model to understand the objective, a world model to predict outcomes, a VLA model to generate actions and a low-level controller to execute those actions safely.

The most important names to remember are NVIDIA GR00T, Physical Intelligence π0, Gemini Robotics, Figure Helix, OpenVLA, SmolVLA, NVIDIA Cosmos, Meta V-JEPA and Google Genie. Together, they represent the emerging intelligence stack for machines that can see, understand and act in the real world.