Runway opens Praxis-1 robot-policy weights and an experimental consumer-GPU path
Runway opens Praxis-1 robot-policy weights and an experimental consumer-GPU path
Runway made Praxis-1 publicly available on November 11, turning its earlier world-action-model announcement into a release robotics teams can inspect and adapt. The company says the weights, inference code and fine-tuning tools are covered by Apache 2.0, including commercial use. Runway announcement
Praxis-1 predicts short sequences of video and robot actions together. The public repository supplies single-arm DROID and two-arm YAM workflows, simulation instructions and LoRA adaptation. The default bf16 quick start calls for an 80GB CUDA GPU. An experimental NVFP4 path targets SM120 cards such as the RTX 5090, but its consumer-GPU peak memory remains unmeasured, and closed-loop validation on SM120 is still absent. Release repository
Runway reports 46.2% success across 1,200 simulated RoboLab-120 episodes. Its technical report also says four leaderboard entries score higher, so the narrower claim about world-action models should not become an overall robotics crown. It identifies problems with similar-looking objects, vague instructions and multi-step tasks. Independent reproduction was not established in this check. Technical report
For developers, the immediate change is access to a trainable robot policy and supporting tools, rather than proof of dependable unattended operation.
Primary sources rechecked November 15, 2026.
Sources and citations
Revision history
Published .