Two robot arms mirroring one pose to lift a shared object: add per-robot offsets, not just a shared pose
Setup: one keyboard controller sends the same gripper-tip pose (turn, reach, height, tilt) to two small arm robots so they lift one long object together; each robot converts the pose with its own calibration and reach correction.
What failed: with a long bar, one gripper closed on the bar and the other closed on air. The arms were fine; the robots simply never stand at exactly the same distance and angle from the object, and 1-2 cm is enough to miss. A shared pose cannot correct that.
What works better: - Keep a per-robot offset (turn / reach / height / tilt) on top of the shared pose. A key selects one robot; while selected, the movement keys change only that robot's offset, so each gripper is lined up on its own end. Then switch back to 'together': every later move applies the shared change and each robot keeps its offset. - Bound the offsets (here 25 deg, 6 cm, 6 cm, 30 deg) and roll back pose AND offsets to the last state every robot confirmed when any robot reports 'out of reach', so the arms never part while holding. - Show each robot's actual solved tilt in the status line (the IK picks the nearest feasible tilt; two robots given the same target settled at -58 and -60 deg), so divergence is visible before gripping. - Once both hold a rigid object, only move vertically; a shared turn or reach moves two gripper points along different arcs and the servos fight. - Any 'go to the fixed parking pose' command ignores offsets: put the object down and open the jaws first.
Network side note: after the robots moved to a new Wi-Fi, one had a new IP and ssh refused it in BatchMode. Instead of blindly accepting, ssh-keyscan the new IP and compare with the key already trusted for that robot's old IP; if it matches, append just that key line to known_hosts.