Runway Releases Open-Weight Praxis-1 World Action Model for Robotics

Runway has officially released Praxis-1, its 27-billion-parameter world action model designed to generate video and control actions simultaneously for physical AI systems. Released with open weights under an Apache 2.0 license, the model allows hardware developers to inspect, adapt, and commercially deploy universal policies across humanoids, industrial arms, and dual-arm warehouse setups, according to The Robot Report.

What happened

According to The Robot Report, Praxis-1 is a mixture-of-experts (MoE) diffusion transformer built on Runway’s video foundation model. By predicting future video frames alongside physical control actions, the model leverages web video in tandem with teleoperated robot data to learn physical dynamics prior to hardware fine-tuning. The model is accessible via GitHub, Hugging Face with native LeRobot support, and vLLM-Omni, with hosted inference provided on Reactor. It runs natively across NVIDIA compute platforms, including data center chips (H100, B200), workstation GPUs (RTX 5090, RTX PRO 6000), and edge hardware like Jetson Thor on physical robots.

Runway tested Praxis-1 across diverse hardware embodiments with early launch partners. Noble Machines deployed the policy on a legged humanoid for heavy industrial handling, where Andy Chen, head of robotics at Runway, told The Robot Report that the model learned a container plank-loading task from 30 minutes of training data, achieving 15 out of 15 successful executions. However, when tasked with clearing an unseen, messy pile of planks into bins, its success rate dropped to six out of 10. Standard Bots tested the policy on a six-axis industrial arm for machine tending and palletizing, while Ultra deployed it on dual-arm warehouse packing stations.

On the RoboLab-120 benchmark, Praxis-1 reached a 46.2% success rate across 1,200 episodes, ranking fifth overall and making it the top-ranked open-source, end-to-end policy on the leaderboard ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%). Fine-tuned on Google's RT-1 robot dataset via SimplerEnv, Praxis-1 achieved an 80.0% success rate, compared to 52.4% for the original RT-1 policy.

Why it matters for robotics and automation leads

For small robotics integrators and factory automation leads, open-weight models with permissive commercial licensing (Apache 2.0) offer a critical alternative to proprietary cloud APIs. The capability to fine-tune a single 27B foundation checkpoint across vastly different kinematics—ranging from a six-axis arm to a dual-arm packer or legged humanoid—using low-rank adaptation (LoRA) reduces the need to build custom control policies from scratch for every specific cell setup.

However, operational teams must separate benchmark performance from production reliability. While Praxis-1 achieved perfect performance on structured plank placement in controlled tests, its success fell to 60% when dealing with unstructured, messy piles. In a live factory or warehouse, a 40% failure rate on variable tasks creates severe line stoppages and requires constant manual intervention. Engineering leads should treat open world action models as strong baseline policies for fast prototyping and domain fine-tuning, rather than drop-in replacements for deterministic motion planners in high-throughput environments.

What to do this week

1. Download the Praxis-1 weights and fine-tuning codebase from GitHub or Hugging Face to evaluate model memory requirements against your edge hardware and workstation setups.

1. Run a low-rank adaptation (LoRA) fine-tuning experiment using 30 to 60 minutes of teleoperated video from a specific, repetitive cell routine like bin picking or machine tending.

1. Audit your target deployment tasks by scene variance, establishing strict manual fallback thresholds for complex or unstructured jobs where policy success rates drop.

Sources

• Runway’s Praxis-1 world action model for robotics is now available — The Robot Report