0.3.2
The observation/action schema is unchanged from 0.3.0.reactor/xwam accepts three camera streams,
a task, and a 16-value end-effector/gripper state. It returns (32, 14) action deltas and
diagnostic predicted states.
0.2.0
Evaluation seed fields became optional. Requests without them use(env_rank, rollout_id, step_id) = (0, 0, chunk_id). Requests with explicit seeds retain their
interpretation.
0.1.0
Initial RoboTwin variant: three camera tracks, 16-value state, and(32, 14) action replies.
The model identifier selects the deployed release. The cookbook follows main; its lockfile records
client dependencies. The simulation guide specifies the
compatible upstream evaluator and RoboTwin revisions.