AIToday
RoboticsOpen-Source AIRobotics & Automation NewsPublished: Nov 23, 2026, 01:00 JST

Runway open-sources Praxis-1 robot model under Apache 2.0 license

Send on LINE
Runway open-sources Praxis-1 robot model under Apache 2.0 license

3 Key Points

  1. What happened

    Runway released Praxis-1, a 27-billion-parameter open-weight world action model, now publicly available under the Apache 2.0 license, with weights and fine-tuning tools on GitHub and Hugging Face.

  2. Why it matters

    The model learns how objects behave from general video footage instead of robot-specific training data, which Runway says could reduce dependence on expensive robot-specific datasets and make advanced manipulation capabilities more accessible to developers.

WHO IT HITSRobotics researchers and developers building manipulation systems, including industrial arms, dual-arm warehouse robots and humanoid platforms, can now adapt Praxis-1 for commercial and research work without licensing fees, though Runway acknowledges it still struggles with visually similar objects and ambiguous instructions.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Praxis-1 was first introduced in September, and its move to a public Apache 2.0 release marks the point where outside researchers and developers can adapt it for commercial and research applications. The design choice at its center is the source of training data: rather than leaning primarily on robot-specific datasets, the model learns how objects behave and interact from general video footage, then is adapted to particular robotic systems. Runway says this could lower dependence on expensive robot-specific training data.

The benchmark picture is mixed. Runway reports a 46.2 percent success rate on Nvidia's RoboLab-120, which evaluates manipulation across 120 simulated tabletop tasks and 1,200 trials, ahead of other open-weight robot foundation models evaluated there, though several closed models and general-purpose AI agents scored higher. On SimplerEnv, Runway reports 80 percent success after fine-tuning for Google's RT-1 robot, compared with 52.4 percent for the original RT-1 policy. Runway also acknowledges limits: the model still struggles with visually similar objects, ambiguous instructions and complex tasks requiring multiple sequential actions, which is roughly where its stated plans for long-term reasoning and autonomous planning point.

FAQ
How does Praxis-1 compare with other open-weight robotics models?
On Nvidia's RoboLab-120 benchmark, Praxis-1 reached a 46.2 percent success rate, versus 42.9 percent for FLUX 3 Action, 36.8 percent for Cosmos 3 Nano and 28 percent for π0.5. Runway notes that several closed models and general-purpose AI agents scored higher.
What robots can Praxis-1 be adapted to?
Runway says the same underlying model can be adapted to industrial arms, dual-arm warehouse robots and humanoid systems, with development partners including Noble Machines, Standard Bots and Ultra.
Where can developers get the model?
The model's weights, inference software and fine-tuning tools are available through GitHub and Hugging Face, with support for the LeRobot framework.
Robotics & Automation NewsRead Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleSamsung, SK hynix, Micron set robot charging standard for chip fabs