The gripper is a joint, not a mode
On these arms the gripper is a Feetech STS3215 on the same bus at the same 7.4 V as the five joints, and it is commanded the same way, as a position. There is no open command and no close command. There is a value, and the policy learns to move that value the way you moved it.
- The gripper value is recorded on every frame, so it is a continuous signal rather than an event the dataset marks.
- The instant that decides the episode, the close around the object, occupies a small fraction of those frames. That underrepresentation is why grasp failures dominate the failure list.
- It is also the most occluded instant, since the fingers sit between a scene camera and the object. A wrist camera is the practical answer.
- There is no force feedback. You command a position and the servo reaches it or stalls trying, and nothing in the observation tells the policy how hard it is squeezing.
What follows for the task you pick
| Question | Answer on these arms |
|---|---|
| Can the policy modulate grip force? | Only indirectly, through the commanded position. There is no force signal to learn from |
| Can I bolt on a different tool? | Mechanically yes, but recorded data belongs to the end-effector it was recorded with |
| Does the gripper need calibration? | It is one of the six actuators the calibration sweep covers, so yes, along with the rest |
| Is a gripper change a retrain? | Yes. Different geometry means a different pixel to action mapping and different reachable grasps |
With no force sensing, a close that begins a moment before the fingers are perfectly placed tends to be more robust than one that waits for the ideal frame, because the object gets funneled into the jaws instead of nudged away by them. Whatever timing you demonstrate is the timing the policy gives back.
You can watch a gripper doing exactly this on real hardware at /live, and /so-100/data-collection covers camera placement for grasp-heavy tasks, where the wrist view stops being optional.