Runway's Praxis-1: 27B open-weight model beats GPT-6 Astra 15.6x at robot control
runwayml · x · 2026-11-20
Runway released Praxis-1, an open-weight world action model that turns its video pretraining into control for real robots, with a technical report.
Architecture & training
- A 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions jointly; the thesis is that pretraining on everyday human video teaches object behavior, hand motion, and task progress transferable to robot control.
Results
- Ranks #1 among world action models and open-source robot foundation models on NVIDIA's RoboLab-120 leaderboard
- With under 30 minutes of teleoperated demos, succeeded in all 15 simulated trials on Noble Machines' robot
- Adapts quickly to new robots: controls a bimanual YAM and a Franka arm on contact-rich everyday tasks (flipping a mug onto a tray, boxing items and closing lids, clearing bottles, closing drawers); fine-tunable on the low-cost SO-101, competitive on Google RT-1 SimplerEnv
vs. LLM-driven arms: on the same station and tasks, Praxis-1 finished 15.6x faster than GPT-6 Astra at 51x lower cost per completed task.
More from Embodied
- Snap a photo on your phone and drop it straight into VR: "spatial airdropping" demo — dreamwieber · 2026-11-20
- Tesla FSD spots deer driver missed at night, sparking regulator critique — AccBalanced · 2026-11-20
- New Google DeepMind robotics scientist shows Gemini controlling robot arm zero-shot — jackyliang42 · 2026-11-20
- RoboEval Challenge winners crowned at CoRL 2026: Ilia tops bimanual manipulation with 0.988 — DJiafei · 2026-11-20
- Walden Robotics CEO Russ Tedrake on sim training and why robot benchmarks are hard — adnothing · 2026-11-20
- Nathan Benaich: frontier robots no longer need custom models per site — shared models improve from every deployment — nathanbenaich · 2026-11-20