World Action Model
also called Robot foundation model
A model that imagines the next few seconds of a scene and turns that imagined future into a robot’s next move.
A model trained largely on video that predicts what happens next in a scene — as video — and derives a robot’s actions from that prediction, instead of mapping camera frames directly to motor commands. It learns how objects behave and how tasks unfold from watching, then adapts to a specific robot with light fine-tuning.
Runway’s Praxis-1 is the clearest example. Introduced in September 2026 and released with open weights on November 11 under Apache 2.0, including for commercial use, it is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions together. On NVIDIA’s RoboLab-120 leaderboard it reaches 46.2% success across 1,200 simulated tabletop episodes, ahead of FLUX 3 Action at 42.9%. The release is unusually complete: three checkpoints (the base model, one fine-tuned on the DROID robot dataset, one fine-tuned for the bimanual YAM robot), inference code, and full and LoRA fine-tuning code on GitHub and Hugging Face with native LeRobot support, running from H100 and B200 data-center cards down to the Jetson Thor computer inside the robot.
The difference from a world model is the loop closing. A world model only simulates — it shows you a place you can move through. A world action model turns the imagined video into motor commands: each call produces 1.6 seconds of motion, 24 action rows at 15 Hz, in roughly half a second on a single H100, according to Runway’s technical write-up. And the caveat belongs next to the headline: 46.2% is first place on a simulated benchmark and still means failing more than half the time, and every figure so far is Runway’s own, with no independent replication published yet. The more interesting number is transfer: fine-tuned on Google’s RT-1 robot, which never appeared in its training data, Praxis-1 reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy.
The pretraining bet is the story: Runway’s research found that web video matches teleoperated robot video as pretraining data for policy performance, which moves the bottleneck from collecting robot demonstrations to how much general video a model can learn from. That is why a video company could ship a robot model at all — the same large-scale video pretraining behind its world models, repurposed for control.
The release is unusually complete for an open model, and that completeness is the point of comparison: three checkpoints, inference code, full and LoRA fine-tuning code, LeRobot support, vLLM-Omni serving, hosted inference on Reactor. Against that, the numbers are still the vendor’s: 46.2% on RoboLab-120 is Runway’s own measurement on NVIDIA’s simulated benchmark, and the SimplerEnv generalization result — 80.0% on a robot it never saw in training — is the more interesting number, because it tests transfer rather than leaderboard position.
The arm that learned from video, not from your lab
Your team has a six-axis arm and a few hundred teleoperated demonstrations, not the thousands a policy normally needs.
Start from Praxis-1’s base checkpoint and LoRA-fine-tune on your demonstrations: the model already knows how objects move from video pretraining, so your data teaches it your embodiment, not physics from scratch. But validate on your own tasks, not on RoboLab-120 — the benchmark is simulated tabletop work, and 46.2% there says nothing about your gripper, your lighting, or your failure modes.
Treating a world action model as a world model, or as a finished robot. It predicts motion, it does not guarantee safe motion — open weights do not make a robot arm safe, and a simulated benchmark is not a factory floor.
- blogPraxis-1 Is Now Available: Open-Weight Robotics Model ↗
The open-weight release: a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, Apache 2.0 including commercial use, 46.2% on NVIDIA’s RoboLab-120. All figures are Runway’s own; the model was introduced in September 2026, the weights opened November 11.
Foundation Model · Base model
A large model trained broadly, meant to be adapted to many downstream tasks.
Multimodal
A model that handles more than text — images, audio, video, screen pixels.
Agent
A model given tools, a goal and permission to keep going on its own.
Open Weights · Open-source model
A model whose trained parameters you can download and run yourself.