Runway Open-Sources Praxis-1 World Action Model, Tops RoboLab-120
On November 11, Runway released Praxis-1, its first world action model, and fully open-sourced it: the weights are available to all developers and researchers under the Apache 2.0 license, runnable and fine-tunable on their own hardware. The core idea is to transfer Runway's massive video pretraining into real robot control capabilities.
Confirmed
- The architecture is a 27B-parameter MoE diffusion transformer that jointly predicts video and actions
- It tops the NVIDIA RoboLab-120 leaderboard with a 46.2% success rate, ahead of π0.5 and GR00T N1.6
- Runway officially states that on the same workstation task, a robotic arm driven by Praxis-1 completed the task 15.6x faster than GPT-6 Astra, at 51x lower cost per task
- Runway believes the most scalable source of robot data is general video: predicting ordinary video teaches the model object motion, hand actions, and intermediate task states, laying the groundwork for switching to new embodiments; after fine-tuning on Google's RT-1 robot (never seen during training), it reached 80% on SimplerEnv, outperforming RT-1
- The model shows strong few-shot adaptation ability (per the official release)
Unconfirmed
- The benchmark results and comparison figures above all come from Runway's official claims, with no independent third-party verification yet; details of the test setup for the GPT-6 Astra comparison were not provided in the materials
Why it matters
- A video generation company directly leveraging its video pretraining into real-world robot control is a landmark case of the "video foundation model → embodied AI" route
- Fully open weights under Apache 2.0 lower the barrier for robotics researchers and developers, potentially putting it in direct competition with existing models like π0.5 and GR00T in the open-source robotics community
2026-11-11 ~ 2026-11-12 · 10 related posts
Primary sources
- [source] Runway open-sources Praxis-1, turning video pretraining into real robot control — runwayml · 2026-11-11
- [source] Praxis-1 tops RoboLab-120 at 46.2%, beating π0.5 and GR00T N1.6 — runwayml · 2026-11-11
- Runway open-sources Praxis-1, a 27B video-action model topping NVIDIA's robot leaderboard — runwayml · 2026-11-11
- Runway: video pretraining lifts Praxis-1 to 80.0% on SimplerEnv vs RT-1's 52.4% — runwayml · 2026-11-11
- Runway's Praxis-1 finishes robot tasks 15.6x faster than GPT-6 at 51x lower cost — runwayml · 2026-11-11
- Runway releases Praxis-1: video-trained world action model tops NVIDIA RoboLab-120 — graceisford · 2026-11-12
4 near-duplicate retellings: runwayml · runwayml · graceisford · jongranskog