Runway’s Praxis-1 bets robot control on video, not teleoperation

Runway, the generative-video company, introduced Praxis-1 in an announcement reported on October 1, 2026: an open-weight “world action model” that turns its video pretraining into control for real robots. The number that matters most is about testing, not control — Runway says that evaluating robot policies inside its world model predicts real-world results with 0.95 correlation, a claim that, if it holds outside Runway’s own setups, would shorten the slowest loop in robot learning.

The weights are not out yet. Runway says Praxis-1 will be released publicly with open weights “in the coming months”. Until then, three hardware companies — Noble Machines, Standard Bots and Ultra — are testing it under early access.

What is a world action model?

Runway describes Praxis-1 as a model that converts video pretraining into robot control, built on the same pretraining behind its general world models and its earlier Solaris and GWM Worlds work. Most robot foundation models in use today come from the other direction: a vision-language backbone fine-tuned to emit motor commands, trained heavily on teleoperated demonstrations (our comparison of the main ones). Runway’s argument is that a model which has already learned from video how objects move, fall and deform starts with an advantage that action data alone does not provide.

Its evidence for the data side of that argument is one published comparison. Pretraining on ordinary web video produced a final placement error of 16.1 cm, against 16.0 cm when pretraining on teleoperated robot video, according to Runway’s announcement. The result shows parity between the two data sources in that experiment; it does not show either is precise enough for fine manipulation. CTO Kamil Sindi framed the bet as a scaling claim: performance “improves as we scale video, so its ceiling is set by how much video it can learn from, not how many robot demonstrations exist.”

Runway says Praxis-1 has run on bimanual arms, a six-axis arm and a mobile base, in studio and home settings, across tasks ranging from lifting soda cans to packing gift bags. It named transparent objects, cluttered scenes, deformable materials such as cloth and rows of near-identical items as target problem cases. It has not disclosed model size or architecture details.

Why does the 0.95 correlation matter?

Real-world evaluation is the bottleneck robot labs rarely advertise. Each policy checkpoint has to be run on hardware many times, with people resetting scenes, before a team knows whether it improved. A world model that ranks checkpoints the way reality would lets a team screen candidates on GPUs and send only the strongest to the robot. Runway says its approach compares favourably with more expensive evaluation based on 3D scene reconstruction; the coverage we reviewed did not give the baseline’s own figure.

Two limits apply. A correlation measures whether predicted and real outcomes rise and fall together across a set of policies; it says nothing about absolute success rates. And the 0.95 comes from Runway’s own tasks and hardware. Independent checks require the weights, which is why the release date matters more than the announcement.

What has to happen before Praxis-1 matters to other labs?

As of October 7, 2026, Runway has given no release date, parameter count or licence. The licence will decide most of the outcome. Permissive terms would let robot makers fine-tune Praxis-1 as a base model on their own hardware; research-only terms would let academics test the evaluation claim but leave commercial users dependent on partner agreements.

For Runway, robotics is a market where its core asset — video pretraining at scale — is the input other labs pay most to approximate through teleoperation. That fits the broader shift of capital toward world models. The claim that video can stand in for much of that teleoperation spend becomes testable once the weights are public.