Hover to animateDataApptronik’s Robot Park makes the fleet part of the training system
Apollo 2’s wheeled and walking configurations collect different physical experience for the Google DeepMind collaboration.
Physical AI
Company releases that turn video, demonstrations and learned models into physical actions—and the environments used to develop them.
Hover to animate12 articles
Hover to animateData
Hover to animateModelsSkild S1 uses video to specify an unfamiliar robot task
The model’s real advance is rapid task specification. Its widely quoted 66% result measures something narrower than autonomous job completion.
Hover to animateModelsDYNA’s next robot has to remember which machine is still running
Dyna-2.1 combines a mobile body, an action policy and a slower workflow model. Its useful memory records what happened in this room today.
Hover to animateModelsRhoda makes the predicted video the robot’s decision
The Direct Video-Action approach separates choosing a future from translating that future into motion. Its small robot-data bill is still not a claim of learning from nothing.
Hover to animateModelsGenesis GENE-26.5 makes the robot hand part of its data strategy
The company’s model, glove and manipulation platform address the translation between human demonstrations and robot actions together.
Hover to animateModelsGeneralist’s GEN-1.5 separates showing a task from retraining a robot
Short demonstrations can act as context, while a separate adaptation route changes weights. Those are different deployment choices.
Hover to animateEvaluationsNVIDIA V2D turns human video into robot practice, one stage at a time
The released workflow reaches GR00T N1.7 training. A small real-robot test shows why simulated practice and real adaptation need to be examined separately.
Hover to animateModelsNVIDIA’s Hydra-0 gives different robots a shared visual action language
Action flow connects human hands and robot arms through image motion. Its promising policy metric measures replay, leaving closed-loop prediction as a separate test.
Hover to animateModelsGoogle’s Genie 3 makes the simulator part of the experiment
Interactive generated worlds could broaden agent training. Their own failures must be separated from the learner’s mistakes.
Hover to animateModelsFigure’s Helix 2.5 makes a trained task portable across unfamiliar homes
A controlled comparison points to human-experience pretraining—not a fresh robot lesson in every room.
Hover to animateModelsPhysical Intelligence makes robot AI easier to access. What should you teach it next?
Physical Intelligence’s beta shifts attention from obtaining a model to knowing exactly what to teach it next.
Hover to animateModelsGemini Robotics 2: Can a new robot reuse another robot's skills?
Google DeepMind’s transfer claims point to a potentially scalable model business: reuse learned skills across different bodies. The test is total adaptation effort, including data collection and validation.
No articles in this subject yet
Choose another subject to continue reading.
What each system does
These releases occupy different parts of the stack. A policy chooses actions; a world model predicts experience; a reconstruction tool prepares demonstrations. Their scores do not measure the same thing.
- Apptronik Apollo: A shared robot fleet connects demonstrations, operational experience and validation in a physical training environment.
- Skild S1: A video demonstration conditions a robot policy without a new training job; per-step scores require different interpretation from uninterrupted task completion.
- DYNA 2.1: Reusable motion and workflow memory work at different timescales, connecting physical movements to jobs that outlast a single action.
- Rhoda DVA: Predicted video supplies the decision, while an embodiment-specific translator converts that prediction into robot actions.
- Genesis GENE-26.5: A manipulation model and its hand hardware form a data strategy. GENE is distinct from the Eno robot and Genesis World simulator.
- Generalist GEN-1.5: Physical demonstrations can prompt a policy or supply gradient updates; the two routes have different adaptation costs and measured results.
- NVIDIA V2D: A video-to-demonstration reconstruction challenge tests which recovered motions help robot learning; it is not itself a general robot controller.
- NVIDIA Hydra-0: Image-space action flow links different bodies. Replaying recorded motion and controlling a robot in a changing scene remain separate tests.
- Google Genie 3: An interactive world model supplies an environment for an agent, rather than the robot policy that acts inside it.
- Figure Helix 2.5: A robot policy tests how far shared pretraining carries a newly demonstrated task across settings.
- Physical Intelligence: An API and demonstration loop connect a general policy to the practical work of collecting useful adaptation data.
- Gemini Robotics 2: A vision-language-action model connects perception and instructions to control, with transfer measured separately from a compelling demonstration.