Runway Launches Praxis-1 Open-Weight Robot AI Model

Runway has introduced Praxis-1, an open-weight AI model built on large-scale video pretraining. The model is designed to translate visual understanding into robot control across diverse hardware embodiments. Runway AI Inc. announced the model this week, describing it as its first 'world action model'. The company plans to ship Praxis-1 with open weights, allowing developers to download and adapt the model themselves rather than accessing it as a closed system.

Kamil Sindi said: "Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect." Runway's approach uses ordinary video to teach models about object behavior and physical interactions. The company states that a policy understanding physical plausibility from video has a significant head start over one built from action data alone. Runway will make Praxis-1 publicly available in the coming months.

Early partners are testing Praxis-1 ahead of public release

Runway is rolling out Praxis-1 to select partners for real-world validation before broader availability. The initial group includes Noble Machines, Standard Bots, and Ultra Robotics. These partners are testing the model on their own hardware platforms. Noble Machines is evaluating it for bimanual manipulation tasks. Standard Bots is running Praxis-1 on its RO1 six-degree-of-freedom robotic arm. Ultra Robotics is conducting tests with a mobile robotic platform. Runway plans to bring on more early-access partners to continue improving the model ahead of its full launch.

Video pretraining enables strong robot policy performance without costly real-world data

Runway claims its video-based approach achieves a 0.95 correlation between simulated robot policies inside its world model and real-world results. The company states this compares favorably with more expensive 3D reconstruction-based techniques. Praxis-1 provides a generalist policy model that works across any robot embodiment or environment. Runway teaches its models to generate accurate physics, hand movements, and task progression for agent training.

The model has been trained on complex tasks involving deformable objects like packing gift bags. Early testing indicates a single model can adapt across very different hardware types, from bimanual arms to humanoids, with light fine-tuning. Runway's experiments show robot policy performance improves with more third-party video used during training. In one experiment, policies pretrained on web video and on teleoperated robot video showed nearly identical final placement errors after fine-tuning.

Runway notes the difference is not statistically significant within reported uncertainty. The company is testing Praxis-1 on manipulation problems in cluttered environments with transparent objects, deformable materials, and groups of similar-looking objects. In a demonstration, the same policy operated in both a controlled studio and a domestic kitchen without retraining. Praxis-1 builds on Runway's interactive video and world models, which represent how environments and objects change over time.

Praxis-1 supports U.S. leadership in physical AI through open access

Runway positions the model as critical to advancing U.S. manufacturing competitiveness. The company believes U.S. leadership in physical AI is essential for regaining a manufacturing lead and that open American models are required to achieve it. Runway views open world models as a compounding advantage that gives hardware developers flexibility and control they currently lack. This move marks Runway's expansion from generative AI video technology into the physical AI and robotics market. The company will continue assessing Praxis-1's efficacy and safety across different embodiments and environments as it onboards more early-access partners.