0.4.2
reactor/cosmos-nano-policy-droid accepts three camera streams, a task, measured joint/gripper
state, and execution acknowledgements. It returns (32, 8) absolute joint/gripper targets.
The first reply needs no execution acknowledgement. Subsequent predictions require an increasing
acknowledged step. Reset restarts the counter and clears camera frames but preserves input state
fields; use a fresh session when old observations or replies could cross an episode boundary.
See the API reference for the contract and the
simulation guide for compatible RoboLab setup.