Runway releases Praxis-1 robot control model with open weights and code under Apache 2.0

Runway has released its robot control model Praxis-1, first shown to select partners in late September, on Hugging Face (after signup and agreeing to access terms) and GitHub under Apache 2.0, which permits commercial use. Versions pre-adapted to specific robots are included alongside the base model, plus a research report with new tests. The base model needs extra training for any new robot, and the standard version needs an 80 GB GPU. Runway reports 46.2% on Nvidia's RoboLab-120 against 42.9% for FLUX 3 Action, 88/100 against 80/100 on real Franka arm tasks, and 80% against RT-1's 52.4% on a new robot type after about 95 hours of fine-tuning.

Robotics researchers and companies can now download, test and build on a video-trained robot controller without a licensing barrier, which lowers the cost of trying the approach. The practical test is whether it really cuts the amount of robot-specific data needed. The 95 hours of fine-tuning for a new robot suggests the effort is reduced, not removed. The 80 GB GPU requirement puts it out of reach of most hobbyists and small teams without rented hardware. Competitors now have a public baseline to compare against.

The "leads World Action Models" claim rests on benchmarks run by Runway itself, with no independent replication yet. The lead over FLUX 3 Action on RoboLab-120 is about 3 points, which may be within normal variation, and 88 vs 80 on 100 real-arm trials is a small sample. Comparing against RT-1, an older model, flatters the 80% result. Whether video pretraining truly reduces robot data needs is still an open question, and the "release" is not new capability, since the model was shown to partners in September.

1. 01Independent replications of the RoboLab-120 and Franka results by outside labs, especially anyone matching or missing the 46.2% and 88/100 figures.
2. 02Third-party fine-tuning reports on new robots: how many hours and how much data they actually need compared with the claimed 95 hours, and on hardware smaller than 80 GB.
3. 03Adoption signals over the next one to three months: Hugging Face downloads, GitHub forks and issues, and named commercial deployments (for example the robot makers listed as partners) using the Apache 2.0 licence.