Runway details how Praxis-1 controls robots
Runway's technical report on Praxis-1 explains how a 27B-parameter video model became an open-weight robot controller: a separate action expert reads the video model's predictions of what happens next.
As we reported, Runway released Praxis-1 as an open-weight world action model. The company has now published a technical report on how it is built. Runway's argument is that everyday video, which people already record in huge volume, can stand in for scarce teleoperation data.
The model has three parts. The first is a video backbone built on Runway's diffusion transformer. It runs in the latent space of a frozen Wan2.2 VAE and uses a frozen Qwen3.5-2B text encoder. It has 26.9B parameters, but only about 2.8B are active per token, because 17 of its 20 feed-forward layers are mixture-of-experts: each has one shared expert, and every token is routed to eight of 128 smaller ones. The second part is a dense, 16-layer action expert with about 314M parameters that reads eight backbone layers through cross-attention. The third is a set of thin per-robot projections that map each robot's poses and actions into a shared format.
The base model generated whole clips at once, so frames could see later frames. A robot can't see the future, so Runway retrained the attention to be chunk-causal: each 1.6-second chunk sees only the chunks before it. On each call the model returns predicted video plus 24 action rows at 15 Hz. Runway also tested placing action tokens inside the backbone. That made no clear difference (+0.01 cm action error) and cost more, so the company kept the separate expert.
According to Runway, Praxis-1 fine-tuned on DROID ranks first among world action models and open-source robot foundation models on the RoboLab-120 leaderboard, and was competitive on a real Franka arm. The excerpt reviewed here ends before the full benchmark tables.