Praxis-1 brings Runway’s video pre-training approach to robotic control.

Runway is extending its video model research into robotics with Praxis-1, a World Action Model designed to control physical robots. The model learns partly from conventional video instead of relying exclusively on robot demonstrations, with open weights planned for the coming months.

Learning before touching a robot

Robot demonstrations are expensive and difficult to collect at scale. Praxis-1 starts with another source of data: the videos humans produce every day.

Runway describes its first World Action Model as a generalist policy that uses video pretraining before being adapted to physical systems. The premise is to learn object behavior, human movement, and how tasks unfold before adapting that knowledge to a specific robot.

The approach extends Runway's work on interactive, real-time video models including Solaris and GWM Worlds 2. Web video gets close to robot demonstrations

Runway compared pretraining on web video with data collected from teleoperated robots. After fine-tuning, final placement error reached 16.1 cm with web video versus 16 cm with robot demonstrations.

The experiment covers 93 evaluation pairs, and Runway notes that differences smaller than the error bars are not statistically significant. In this test, the two pretraining sources therefore produced similar performance.

The company also reports that policy performance improves as the amount of third-person video increases. That relationship is central to Praxis-1: shifting part of the robotics data problem toward a source available at much larger scale. Simulated environments to evaluate policies

Runway's World Model research also plays a role in evaluation. The company says it compared policies simulated inside its World Model with their performance on physical robots.

Runway reports a 0.95 correlation between the two, which it says compares favorably with more expensive 3D reconstruction-based techniques. The result comes from Runway's own research and will need broader external evaluation across additional hardware configurations.

Praxis-1 is also demonstrated on multi-step tasks. A mobile robot approaches a shelf, positions itself, picks up a book, and removes it from the environment. Other tests focus on situations that can challenge policies trained only on demonstrations, including transparent, deformable, occluded, and visually repetitive objects. One policy across different robots

Runway describes Praxis-1 as independent of a specific embodiment or environment. Early testing is underway with Noble Machines, Standard Bots, and Ultra, each running the model on its own hardware.

The configurations include bimanual manipulation, a six-degree-of-freedom robotic arm, and a mobile base. Another demonstration moves the same policy from a controlled studio environment to a domestic kitchen without retraining.

This early phase is intended to identify performance and safety gaps across different hardware and environments before public release. Open weights coming in the next few months

Runway plans to release Praxis-1 with open weights rather than keeping the model behind a closed API. The weights are not available at the time of the announcement.

The company connects that decision to interoperability in robotics, where developers need flexibility to adapt models to different hardware and control systems.

An early-access program is open to teams interested in testing Praxis-1 on their own robots. The open weights are expected in the coming months, with no more specific release date provided.