Introducing Praxis-1

TL;DR

Runway’s first announced World Action Model, Praxis-1, is a generalist open-weight robot policy built on the same large-scale video pretraining that backs their general world models (Solaris, GWM Worlds 2). It is positioned as a cross-embodiment / cross-environment policy that treats video — including generated video — as the dominant data substrate rather than teleoperated demonstrations. Currently being tested with Noble Machines, Standard Bots, and Ultra ahead of a planned open-weight public release in the coming months. No architecture, parameter count, action-head details, or benchmark numbers are disclosed; the post is launch-signal rather than technical.

Key claims

Praxis-1 is Runway’s first open-weight world action model, built on the same large-scale video pretraining behind their general world models [§1].
The bet is that video is effectively limitless and becoming infinite as generative models reach parity with real video on quality and speed, which makes it the right pretraining substrate for generalist robot policies — scarce teleoperated data cannot cover edge-case-heavy deployments like autonomous driving or household robotics at the required volume [§2].
Policy performance improves as third-person web video is scaled, and web video matches teleop robot video on at least one measurement axis (Figure 2 caption) — the bottleneck on a capable policy becomes how much general video a model can train on [§3, Fig. 2].
Praxis-1 is being tested with Noble Machines, Standard Bots, and Ultra on each partner’s own hardware ahead of public release, with the testing explicitly framed around identifying gaps across embodiments and environments [§4].
Runway’s prior work on simulating robot policies inside their world model is reported to predict real-world results at r = 0.95, compared favorably to more expensive 3D-reconstruction-based evaluation techniques [§2, cross-reference to Runway’s accelerating-robot-policy-evaluation post].
Open-weight release is pitched as a strategic bet on U.S. physical-AI leadership — Runway argues interoperability and openness are the compounding advantage hardware developers currently lack [§5].

Method

Not disclosed. The post only states that Praxis-1 is built on Runway’s existing video-pretraining stack (shared backbone lineage with the Solaris and GWM Worlds 2 interactive world models), that it is a generalist policy expected to work across embodiments and environments, and that the deployment path uses partner hardware. There is no architecture description, no parameter count, no action-head details, no training data composition, no fine-tuning recipe, and no information on how the video pretraining is adapted into action-conditioned control. Figures (multi-step tasks in complex environments, web video matches teleop robot video, where policies usually break, studio/home) are illustrated with video captions only.

Results

None disclosed. The post reports a qualitative “policy performance improves as we scale third-person video” scaling claim with a figure caption but no numbers, task definitions, or baselines. The r = 0.95 real-vs-simulated correlation is a prior Runway result for policy evaluation inside their world model, not a Praxis-1 capability result. No head-to-head against π0.5 / GR00T / Cosmos 3 policies / Dyna-2 / Isaac 0.5 / Spirit-v1.5 / Embodied-R1.5 / 1XWM. No per-task success rates, no latency numbers, no benchmark scores on LIBERO / RoboTwin / RoboArena / RoboLab / VANTAGE-Bench.

Why it’s interesting

This is the “GWM Robotics” variant from Runway’s December 2025 GWM-1 launch (Runway GWM-1 launch broadcast — world model, avatars, audio, multi-shot editing) finally shipping as its own named product with explicit open-weight commitment, which turns the Runway/Odyssey/Cosmos closed-vs-open framing more interesting. Praxis-1 extends the “video-model-as-policy” thesis the wiki has been tracking — most sharply in Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners (Rhoda AI’s DVA: causal web-video pretrained model used as the policy at every control step) and 1XWM: From Video to Action — 1X's video-pretrained world model as a NEO robot policy (1X’s 1XWM-as-policy on NEO) — but adds the first filed instance from a closed-flagship video-generation lab publicly committing to open weights on the robotics side rather than keeping the WFM/WAM stack closed as DeepMind (Genie 3), Odyssey (Odyssey-3), and Waymo (Genie-3 driving) have done.

Praxis-1 is also the first open-weight entry to pair the World Foundation Models “pretrain on video, deploy as policy” position with the explicit cross-embodiment-partner release pattern that ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining / Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training / Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) have been pursuing at smaller scale. Sibling framing to Cosmos 3: Omnimodal World Models for Physical AI (NVIDIA Cosmos 3 open omnimodel including VLA/world-action) but on a video-generation-first rather than omnimodel-first substrate. The r = 0.95 real-vs-sim correlation claim, if replicable at Praxis-1 scale, would also sharpen the RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation / Evaluating Gemini Robotics Policies in a Veo World Simulator / SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation WFM-as-policy-evaluator thread.

Treat as launch-signal / market-structure datapoint until weights land. All substantive technical questions — architecture, action space, pretraining-vs-finetuning split, how video-only pretraining is adapted into closed-loop control, actual scaling curves — are deferred to the public release.