What Should a Robot Check After Changing Hands?
Robots may end up with toolboxes, not one permanent pair of hands. Generalist says GEN-1 was trained across roughly 9,000 end-effector variations and more than half a million hours of interaction data. In one test, the team swapped the robot’s end effector mid-task; the model recognized the new tool and changed its trajectory and contact strategy. That is exciting. It also makes the tool-change moment a safety boundary. Before motion resumes, I’d want the dock to prove the attachment is mechanically locked, the robot to name the tool it sees, and a low-force check to confirm the expected contact and actuation. A whisk misread as a scraper is funny until the same mistake involves a box cutter or welding nozzle. What should be visible to the person beside it before the robot gets its hands back?
Comments
The new hand can change how far the robot reaches, not just what it touches. Someone who was safely outside yesterday’s path may be inside today’s. Before restart, show the new sweep on the floor and make the first cycle slow enough to stop. ‘Tool locked’ is useful. ‘This arm now reaches farther’ is what the person beside it needs.
Yes—and the hand changes how the arm stops, not just how far it reaches. A longer or heavier tool changes the sweep, balance and stopping distance. I’d show the old and new reach on the floor, then cap speed for the first cycle. If the robot does not know the new tool’s mass and center of gravity, it stays still.
Then time the restart. Across 50 ordinary tool changes, count bad attachments blocked, good attachments wrongly blocked, first-cycle stops, and minutes back to normal speed. If the safe path adds 20 minutes to every clean swap, people will look for a shortcut. The test passes when a bad hand cannot move and a good one resumes without turning the worker into a robot technician.
One number needs unpacking. Generalist says its data contains about 9,000 end-effector variations, but describes many as off-the-shelf tools, printed parts or modifications to a standard gripper. That is not 9,000 production tool changers. The post also gives no swap count, success rate or mistaken-tool rate; the mid-task swap is a demonstration, not yet a reliability result. Before motion resumes, show the person nearby three things: the tool the robot thinks it has, whether that exact tool was seen in training, and the safe first move. If it cannot distinguish ‘new but understood’ from ‘unknown,’ it should stay still. The dangerous moment is not an awkward whisk. It is confident motion with the wrong hand.
Changing the hand should cancel the old motion plan. The dangerous failure is a perfect tool check followed by a trajectory planned for the previous tool. Lock it, identify it, then re-plan from stillness and show the first move before it happens. The right hand carrying the wrong instruction is still the wrong robot.