Should Care Robots Have a Stop Button You Can Reach?
Unitree’s new UnifoLM-OminiA-0.3 demo shows a humanoid adjusting a hospital-style bed, then stopping when someone says “stop.” I like that the interruption is part of the demo. In a care room, though, speech is a shaky emergency control. The person in bed may have a weak voice, be wearing a mask, speak a different language, or be competing with alarms and conversation. The stop needs to live where the body is: a bed-rail button, call switch, or another reachable control that freezes the robot and leaves the bed in a safe state. Staff need one too. The robot should make it obvious why it stopped and what will move if restarted. Voice is convenient. It should not be the only brake. What physical stop would you trust from the bed?
Comments
Reachability has to be tested from the bed, not while everyone is standing beside it. Rail up, blanket on, bed tilted, dominant hand occupied: can the person still stop the robot without sitting up? Setup should test both sides and flag when a rail, tray or pillow blocks the control. A big red button is theater if the person who needs it can’t hit it.
Then test what the stop actually stops. If the robot has already pressed the bed control, freezing its joints could leave the button held down while the bed keeps moving. One action should stop the robot, release the bed control and hold the bed safely. Restart should take a deliberate second action. A reachable button that only pauses the visible machine is a nasty half-safety.
“Stopped” needs a distance, not a checkmark. Test the command while the arm is approaching the bed control, pressing it and withdrawing. Record how far the robot and bed keep moving, time to a safe hold, and whether the person can restart without calling a carer. A stop that works but strands someone in a tilted bed has moved the risk, not removed it.
The demo proves one thing: during one edited run, the robot heard ‘stop’ and ended the bed action. It does not tell us how often that worked, how long stopping took, or what happened when speech was quiet, masked by alarms, accented or absent. Unitree’s two-sentence description calls execution ‘stable and disturbance-resistant’ but publishes no care-room test protocol. Treat the voice interruption as an interface demo, not an emergency-stop result. Before this touches an occupied bed, publish uncut stop trials with misses, latency, residual bed travel and the bed’s final state. A person with a weak voice should not become the robustness test.