Runway opens Praxis-1 robot model under Apache 2.0; its own tests put the weak spot in the video

Runway has released Praxis-1, the robot foundation model it first showed in September, with open weights under the Apache 2.0 license, which allows commercial use. The 27-billion-parameter mixture-of-experts diffusion transformer is a world action model: like a world model, it predicts video, and it predicts robot actions alongside. It went up on GitHub and Hugging Face on Nov. 11 with its inference and fine-tuning code, and Runway says it scores 46.2% on NVIDIA's RoboLab-120 simulation benchmark, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%.

Runway's own technical report puts the model's weak spot in its imagination, not its motor skills. Given the true future video of a Franka arm, on 32 held-out DROID clips, Praxis-1's predicted end-effector positions were off by 1.14 cm. Made to imagine that future itself, as it must on a real robot, the error rose to 5.71 cm, five times as much, and Runway attributes most of it to the imagined video. If that holds, better control in this design comes from better video prediction, which is what Runway, a video generation company, already sells.

Only 18% of the training video shows a robot

Runway built the model on a backbone pretrained on general video. In the next training stage, egocentric human video made up 58% of tokens, general video 20% and real-robot video 18%. Fine-tuned on Google's RT-1 robot, which was not in its training data, Praxis-1 reached 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy. On a real Franka arm it succeeded in 88 of 100 episodes, against 80 for FLUX 3 Action.

Launch partners Noble Machines, Standard Bots and Ultra run it on a legged humanoid, a six-axis arm and a dual-arm warehouse station. Andy Chen, Runway's head of robotics, told The Robot Report that with Noble Machines, Praxis-1 learned to pick up wooden planks from half an hour of examples and got it right 15 times out of 15. On a messy pile it had never seen, it cleared the whole pile six times out of 10.

Agent systems still sit above it

Runway's "top-ranked" claim holds among world action models and open end-to-end policies. On NVIDIA's leaderboard Praxis-1 is fifth overall. AMP1, an agentic language-model system, scores 86.0%, and SenseNova-RoboRSI, also agentic, 57.1%. With the weights open, any lab can now test whether fine-tuning a video-first policy closes a gap of almost 40 points.