Runway, best known for generative video, released its robot model Praxis-1 with open weights and source code on Saturday under the Apache 2.0 licence, commercial use included. It is the clearest attempt yet by a video company to turn an understanding of moving images into a robot controller — and the first that anyone can download.
The model was announced in late September, but access was limited to selected partners. That phase is over: weights and code are on Hugging Face and GitHub, behind a sign-up and acceptance of terms. Pre-adapted versions for different robots ship alongside the base model, which still needs additional training before it can drive a new machine.
The architecture follows what Runway calls a World Action Model. Praxis-1 takes camera images, a text instruction and information about the robot's pose, predicts how the scene will change, and derives the movements that should follow. It corrects itself as new frames arrive. Training ran in stages: a video model that learned from ordinary footage how objects move and interact, then fine-tuning on recordings of human actions and real robot data.
The appeal is in the data bill. Robot datasets are expensive because they are usually collected specifically for the purpose. Runway used about 277 hours of the public DROID dataset to adapt Praxis-1 to a Franka robot arm, and between 1,400 and 2,600 hours for a dual-arm robot. Whether the approach genuinely needs less robot data overall than comparable models without video pretraining remains open; the figures do not settle it.
The performance numbers come from Runway's own tests. On Nvidia's RoboLab-120 benchmark, Praxis-1 scores a 46.2% success rate, leading the World Action Model category; FLUX 3 Action, built on the same premise, manages 42.9% under the same conditions. Overall it ranks fifth, behind two AI agents and two robot models of different architecture. On a real Franka arm it completed 88 of 100 attempts across ten tasks, against 80 for FLUX 3 Action. Transferred to a robot platform it had never seen — the one Google used to develop RT-1 — it reached 80% after about 95 hours of fine-tuning, where RT-1 scored 52.4% on the same benchmark. Runway ran every test itself; none has been independently verified.
The limitations are in the research report as well. Praxis-1 sometimes confuses similar-looking objects, copes poorly with vague instructions, and cannot reliably plan long action sequences — individual steps have to be spelled out. Runway positions it as an execution layer that will eventually work beside general-purpose models that break tasks into steps. The standard version needs a GPU with 80 GB of memory; a compressed build should fit on a 32 GB RTX 5090 according to Runway's estimates, though the docs concede peak memory has not been measured.
Standard Bots, Noble Machines and Ultra have been testing the model since September on a humanoid, an industrial arm and a dual-arm warehouse robot. That those results can now be reproduced is the point: an open model that tops a public benchmark can be checked — and contradicted — by any lab with a GPU.

