Runway releases Praxis-1, an open robot-control model under Apache 2.0

Runway has released Praxis-1, a robot-control model it calls a “world action model”, with open weights under the Apache 2.0 licence, ready to run and fine-tune on your own hardware. The model is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, and Runway claims it ranks first on NVIDIA’s RoboLab-120 leaderboard.

The release on 11 November follows Runway’s first announcement on 30 September, when the company opened early access to a small group of robotics partners and said the weights would come “in the coming months”. Runway is best known for generative video, and Praxis-1 is its first step into robotics.

What Runway released on 11 November

Runway says it is “making Praxis-1 available to everyone”, and that the model “turns video pretraining into control for real robots”. In robotics terms, Praxis-1 is a policy model: the component that decides what a robot does next from what it sees. Runway describes it as a policy model for robotics developers and researchers that works “across any embodiment or environment”. “Embodiment” here means a particular robot body, such as a two-armed station, a single industrial arm or a mobile base.

On architecture, Runway’s description is short. Praxis-1 is a 27B-parameter mixture-of-experts (MoE) diffusion transformer that predicts video and actions together. A mixture-of-experts model routes each input through a subset of specialised sub-networks rather than the whole model. A diffusion transformer generates its output by refining noise step by step, the same broad family of techniques used in video generators. Runway says the architecture, training stages and results are detailed in a technical report, but the report’s contents are not part of the launch posts.

Spec	Praxis-1 (as Runway states it)
Model type	World action model / robot policy
Parameters	27B, mixture-of-experts
Architecture	Diffusion transformer predicting video and actions together
Licence	Apache 2.0, open weights
Deployment	Run and fine-tune on your own hardware
Early partners	Noble Machines, Standard Bots, Ultra

Why Runway is training robots on video

The central bet behind Praxis-1 is about data. In September Runway put it plainly: “Robot demonstration data is limited. But video data is infinite.” Collecting real-world robot data is slow and costly, especially when people have to teleoperate (remotely drive) robots through tasks again and again.

Runway argues that a model trained to predict general video “learns how objects behave, how hands move and what a task looks like partway through”, which gives it a head start on new robot bodies. In earlier comments reported by Robotics & Automation News, the company said:

“A policy that already understands physical plausibility and object behavior from video pretraining has an enormous head start on one built from action data alone.”

Runway compares this with how large language models learn language broadly from text before being adapted to narrower tasks, and says its experiments show a similar scaling effect, with robot policy performance improving as more third-person video is used in training. In one experiment it reported, a policy pretrained on ordinary web video reached a final placement error of 16.1 cm after fine-tuning, against 16.0 cm for one pretrained on teleoperated robot video, a difference Runway itself says is not statistically significant. Put simply, Runway’s claim is that cheap web video got about as far as expensive robot footage in that test.

Praxis-1 also builds on Runway’s interactive video and world models, including Solaris and GWM Worlds 2. The company has said that simulating robot policies inside its world model produced results with a 0.95 correlation to later real-world performance.

Runway’s benchmark and speed claims

All the performance numbers so far come from Runway. None has yet been reproduced by an outside tester in the sources available.

RoboLab-120 leaderboard

Runway calls Praxis-1 “the strongest world action model available” and says it is “#1 on NVIDIA’s RoboLab-120 leaderboard, and ahead of every open-weight VLA on the board, including π0.5 and GR00T N1.6”. A VLA, or vision-language-action model, is a robot policy that takes in camera images and instructions and outputs actions. Runway’s post does not give the scores or margins, so the size of the lead is not known from the announcement.

Versus a frontier LLM driving the arm

Runway also compares Praxis-1 with a general-purpose large language model used as a robot controller. On the same station and tasks, Runway says, Praxis-1 “finished 15.6x faster than GPT-6 Astra driving the arm, at 51x lower cost per completed task”. The post does not describe the tasks, the hardware or how cost was calculated.

Adapting to new robots and new tasks

• Unseen robot: fine-tuned on Google’s RT-1, a robot Praxis-1 never saw in training, the model reaches 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy, in Runway’s evaluation. SimplerEnv is a simulated evaluation suite for robot manipulation.
• Little data: with under 30 minutes of teleoperated demonstrations, Runway says Praxis-1 succeeded in all 15 simulated trials on Noble Machines’ robot. That is a small trial count, and the trials were simulated rather than on physical hardware.

Who tested Praxis-1 before release

Runway ran an early-access programme with three robotics companies: Noble Machines, Standard Bots and Ultra. According to Robotics & Automation News, Noble Machines evaluated the model for bimanual (two-armed) manipulation, Standard Bots ran it on its RO1 six-degree-of-freedom robotic arm, and Ultra tested it on a mobile robot platform. The programme was meant to evaluate both performance and safety across different robot set-ups and environments.

Runway has also shown Praxis-1 on harder manipulation problems, including cluttered scenes, transparent objects, deformable materials and groups of similar-looking objects, and has demonstrated the same policy running in a studio and a domestic kitchen without retraining. These are Runway demonstrations, not independent evaluations.

What open Apache 2.0 weights mean for builders

Apache 2.0 is a permissive licence that generally allows commercial use, modification and redistribution. Open weights mean developers can download the model and run or adapt it themselves rather than only reaching it through Runway’s servers. Runway frames this as a strategic choice tied to interoperability between AI models and robot hardware:

“We view open world models as a compounding advantage that gives hardware developers flexibility and control they don’t currently have.”

For teams building robots, the practical appeal is a pretrained starting point that Runway says can be fine-tuned to a new robot body or task with relatively little demonstration data. For teams that have been prototyping robot control with general-purpose LLMs, Runway’s speed and cost comparison is aimed squarely at them, though the claim needs independent checking before it can guide architecture decisions.

Praxis-1 sits alongside other recent moves in robot foundation models, including DeepMind’s Gemini Robotics 2.5 and its open robot safety benchmark.

Limits and open questions

• Self-reported results. The RoboLab-120 ranking, the GPT-6 Astra comparison, the SimplerEnv score and the Noble Machines trials all come from Runway’s own posts. No third-party reproduction is available in the sources yet.
• Thin public detail. Runway’s launch posts point to a technical report for architecture, training stages and results, but they do not state the RoboLab-120 scores, the comparison tasks, or how cost per task was measured.
• Hardware needs. Runway says the model runs on your own hardware but its posts do not say what compute a 27B MoE policy needs for real-time control.
• Simulation versus reality. Several headline results are from simulation, including SimplerEnv and the 15 Noble Machines trials. Runway’s reported 0.95 correlation between world-model simulation and real performance comes from earlier experiments.
• Safety. The early-access programme was designed partly to assess safety across robot configurations, but no safety findings are given in the launch posts.
• Training data. The sources do not say what video Praxis-1 was pretrained on or whether training code or data are released alongside the weights.

What to watch next

The first signal will be independent results: whether outside teams can reproduce Runway’s RoboLab-120 position and its SimplerEnv score once they run the open weights themselves. The second is the technical report, which should show how the model was trained and how the speed and cost comparison with GPT-6 Astra was set up. Finally, watch for fine-tuned versions from robot makers beyond the three early partners, which would test Runway’s central claim that a model pretrained on general video can transfer to almost any robot body.