Runway open-sources 27.2B Praxis-1 world action model under Apache 2.0 for robot control

Runway open-sourced Praxis-1 under the Apache 2.0 licence on 11 November 2026, releasing base and embodiment-tuned checkpoints on Hugging Face alongside reference training and inference code on GitHub (runwayml/praxis-1-oss), a day after Google DeepMind unveiled Gemini Robotics 2.5. Built to overcome the scarcity of teleoperated robot trajectories by transferring physical intuition from general human video pretraining, Praxis-1 pairs a 27.2B-parameter Mixture-of-Experts chunk-causal video-diffusion trunk (54.4 GB in BF16, using frozen Wan2.2 VAE and Qwen3.5-2B encoders) with a 0.3B-parameter action expert mid-trained across 18 robot embodiments.

Joint video-action chunking and RoboLab-120 results

On each control call, Praxis-1 ingests a language instruction, recent camera frames, and current proprioceptive pose to jointly predict a future video chunk and 1.6 seconds of motor commands (24 action steps at 15 Hz), adapting across single-arm Franka DROID rigs, bimanual YAM arms, Google RT-1 SimplerEnv, and low-cost SO-101 hardware via six lightweight per-rig projection tensors and NVFP4 quantised checkpoints. Runway’s benchmark charts on the research page are interactive web components rather than static image files: fine-tuned on DROID, Praxis-1 leads all world action models and open-weight vision-language-action policies on NVIDIA’s 120-task RoboLab-120 simulation suite, ahead of FLUX 3 Action, pi 0.5, and GR00T N1.6, trailing only zero-shot frontier VLM harnesses AMP1 (86.0%) and SenseNova-RoboRSI (57.1%).

Key numbers

• 46.2% success rate (555 of 1,200 episodes) on NVIDIA RoboLab-120, ahead of FLUX 3 Action at 42.9%
• 27.2B MoE video-diffusion trunk plus 0.3B action expert (54.4 GB BF16 across 17 safetensors shards, plus NVFP4 variants)
• 24 action steps at 15 Hz (1.6 seconds of motion) predicted per chunk alongside future scene video