Available variants
Praxis-1 is a video-action policy from Runway. For each call, it predicts the next chunk of video
and the next chunk of robot commands. Its inputs are recent camera frames, the measured end-effector
pose, and a task instruction. Reactor returns only the commands. The predicted video stays on the
server.
Each variant is a separate checkpoint that was trained for one robot rig. All variants use the same
session protocol. Their cameras, robot state, and command fields are different, so a client for one
variant does not work with the other.
Start with your rig’s overview: DROID,
YAM. Then get your first
actions and read the
reference for the wire contract.