What are world models
and VLAs?
Two kinds of AI model are reshaping robotics. A world model is a neural network that predicts what the world will look like next, given an action, so it can act as a learned simulator. A VLA, or vision-language-action model, is a neural network that takes camera images and a text instruction and outputs robot movements directly.
World models
A world model is trained on huge amounts of video, sometimes with the actions that produced it. It learns to answer "if this happens, what comes next?" Meta's V-JEPA 2 (June 2025) was trained on more than a million hours of internet video and, with a little robot data, could plan manipulation tasks it had never been shown.[4]
The models are becoming generators of worlds. Google DeepMind's Genie 3 (August 2025) produces interactive 3D environments from a text prompt at 24 frames per second.[3] NVIDIA's Cosmos platform, first released in January 2025, offers pre-trained "world foundation models" that developers adapt to generate training data for robots.[1] Cosmos 3 (2026) merges vision reasoning, world simulation and action prediction into a single model.[2]
In robot learning, a world model can replace part of the physics simulator: the robot practises inside the model's predictions. An April 2026 survey traces this idea from early imagination-based RL to today's foundation-model simulators.[15] The limitation is the same as for any learned model: it reproduces the physics present in its training videos. Heat inside a motor, current in a battery or the pull of a magnet do not show up on camera, so a video-trained model cannot predict them. This is the gap between Physical AI and physics that Invisible Force bridges.
Vision-language-action models
A VLA starts from a vision-language model, the kind of AI that can describe a picture, and adds an "action head" that outputs motor commands. The result is one network that looks at the camera, reads "put the cup in the sink", and moves the arm.
The reference examples are from 2024: OpenVLA, an open model trained on about a million recorded robot demonstrations,[14] and π0 from Physical Intelligence, which controls several different robot types from one policy.[5] Since then the pace has been fast:
- π0.5 (April 2025) cleans kitchens in homes it has never seen; π*0.6 (November 2025) improves from its own deployment data; π0.7 (April 2026) recombines learned skills for new tasks without extra training.[6][7][8]
- Gemini Robotics (March 2025) added a planner that "thinks before acting" in version 1.5 (September 2025) and extended to whole-body humanoid control in Gemini Robotics 2 (July 2026).[9][10][11]
- GR00T N1 (NVIDIA, March 2025) is an open VLA for humanoids trained on real, human and synthetic data; N1.6 (September 2025) adds reasoning for ambiguous instructions.[12][13]
Why physics still matters
A VLA maps pixels straight to motor commands, with little in between. It has no notion of motor temperature, battery charge or holding force unless something in its training data encoded them. Physics-based simulation gives these models their training worlds and, at deployment, a safety guardrail. See What is the sim-to-real gap?
References
- NVIDIA, arXiv 2501.03575. Cosmos World Foundation Model Platform for Physical AI. January 2025.
- NVIDIA Newsroom. NVIDIA launches Cosmos 3, the open frontier foundation model for physical AI. 2026.
- Google DeepMind. Genie 3: A new frontier for world models. August 2025.
- Meta AI, arXiv 2506.09985. V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. June 2025.
- Physical Intelligence, arXiv 2410.24164. π0: A Vision-Language-Action Flow Model for General Robot Control. October 2024.
- Physical Intelligence, arXiv 2504.16054. π0.5: a Vision-Language-Action Model with Open-World Generalization. April 2025.
- Physical Intelligence. π*0.6: a VLA that Learns from Experience. November 2025.
- Physical Intelligence. π0.7: a Steerable Model with Emergent Capabilities. April 2026.
- Google DeepMind, arXiv 2503.20020. Gemini Robotics: Bringing AI into the Physical World. March 2025.
- Google DeepMind. Gemini Robotics 1.5 brings AI agents into the physical world. September 2025.
- Google DeepMind. Gemini Robotics 2 brings whole body intelligence to robots. July 2026.
- NVIDIA, arXiv 2503.14734. GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. March 2025.
- NVIDIA Newsroom. NVIDIA accelerates robotics research and development with new open models and simulation libraries (GR00T N1.6). September 2025.
- Kim et al., arXiv 2406.09246. OpenVLA: An Open-Source Vision-Language-Action Model. June 2024.
- Hou et al., arXiv 2605.00080. World Model for Robot Learning: A Comprehensive Survey. April 2026.
Every link was opened and checked in September 2026. Links open in a new tab.