Praxis-1: Runway’s Robot Policy and Early Access

A robot policy must decide what movement to execute from its observations and task. Praxis-1 starts from Runway’s large-scale video pretraining: everyday footage supplies examples of object behavior, hands and intermediate task states. Robot-specific adaptation connects that visual prior to the controls of an actual machine.

For a robotics team, the useful question is how much demonstration collection this removes from a new deployment. A policy that recognizes an object still needs the correct camera inputs, action representation and control timing. Begin with one embodiment and one measurable manipulation task before expanding to a different robot.

Runway’s research page shows rigid repeated objects, cluttered tableware, transparent materials and deformable cloth. It also shows a policy moving from a studio to a kitchen without retraining. These are company demonstrations of the conditions the team is targeting.

The early partners are Noble Machines for bimanual manipulation, Standard Bots with its RO1 arm, and Ultra with a mobile base. Use these configurations to discuss hardware compatibility and the target task when requesting partner access.

In Runway’s pretraining comparison, final placement error after fine-tuning is 16.1 cm for web video and 16.0 cm for teleoperated robot video. The figure uses 93 evaluation pairs and ±1 standard error of the mean; Runway says differences smaller than the error bar are not significant. The nearly equal placement errors make web video worth evaluating when robot demonstration collection is the bottleneck.

As of October 3, 2026, the entry point is Runway’s robotics team for pre-launch partner access. The public release is planned with open weights, but there is no downloadable Praxis-1 checkpoint or published license to install from this announcement.

Describe your robot, camera arrangement, controller, target task and available demonstrations when contacting the team. Ask for the inference hardware, action-space documentation, control frequency and adaptation procedure needed for your setup. A Runway creative subscription or video API key is not an access route to this robot policy.

Griffin Alpha-S offers downloadable base weights and separate LIBERO and RoboTwin fine-tunes. FLUX 3 Action has its own robot checkpoints and LeRobot integration. Those releases are more direct starting points if your immediate goal is to inspect code and reproduce a control loop. Praxis-1 is a candidate for teams that can participate in Runway’s partner testing.

Compare policies on the same robot, camera setup and demonstration budget. Measure task completion, placement accuracy and recovery time separately, then check adaptation to the controller you intend to use.

Physical Intelligence’s π0 is also a generalist robot policy. Its published design starts from a pretrained vision-language model and trains across robot datasets, then specializes with higher-quality demonstrations. Praxis emphasizes ordinary-video pretraining as a foundation for robot adaptation. Choose a training route around the demonstrations you can collect and the controls your robot supports.

Cosmos is a broader platform of world models and video-processing tools. NVIDIA describes using it with Isaac Sim to augment synthetic training data. That pipeline supplies environments and training data for building and evaluating a robot policy. When comparing products, identify whether you need a policy to execute actions, a data-generation tool, or a simulation environment.

Runway has not published Praxis-1 inference prices, parameter counts or GPU requirements. Build a trial budget around the robot and cameras, demonstration collection, adaptation work, inference hardware and operator time. Ask Runway for the partner trial’s hardware and support costs.

Use a held-out set of objects and scenes, log failed grasps and recovery attempts, and measure time per completed task. These measurements let you compare a partner trial with an available open-weight policy even before there is a standard per-call price.