0.1.0
reactor/praxis-1-yam accepts three camera streams: top, wrist_left, and wrist_right. It also
accepts a task, both arms’ measured end-effector pose and gripper state, and an executed-step echo.
It returns absolute end-effector pose and gripper targets: 21 rows for the first chunk of an episode
and 24 after.
The first chunk needs no echo. Each later chunk waits for the last step. reset_context starts a
new episode without resetting the step counter.
See the API reference for the contract.
Known limitations
- Validated in simulation only. Reactor has not yet run this release end to end with a physical robot.
- Only 16:10 camera views are supported. The training data also has 16:9 and 4:3 YAM cameras, which this release does not cover.