Lesson 03 · 15 minutes · Preparation concept

A grasp is a physical
relationship.

Your next win: explain why a convincing animation is not enough to call an object grasped or handed off.

This lesson prepares you to inspect the future SO-101 contact code. That code and scene do not exist yet. The pendulum lab does not test a grasp.

Contact is a constraint, not a magic attachment

When a finger touches a spoon, the simulator computes a contact force that prevents interpenetration. Friction can resist sliding along the contact surface. A successful grasp needs enough geometry, normal force, friction, and a feasible approach for the object to stay controlled when the arm moves.

Teleporting an object into a gripper or welding it to a hand would hide exactly the failure we need to learn from. The planned evaluator will therefore check contacts and object stability after release, rather than trust an animation.

A cup slips while a gripper appears closed. What is the best first hypothesis?

A hand-off has an overlap, then a transfer

Donor holds
object stable
→Receiver closes
contact established
→Verify overlap
both contact
→Donor releases
receiver retains

The arms share workspace during the overlap. A supervisor assigns ownership so both policies do not independently pull the object in incompatible directions. The donor must not release simply because the receiver reached a pose; the receiver needs a stable physical hold first.

Which signal should permit the donor to release?

Practice the inspection question

Imagine the future log says “hand-off complete,” but the spoon falls after the donor opens. Name the three questions you would ask first: about contact, ownership, and the post-release observation. Share your answer with the agent for feedback.

Primary reading: MuJoCo’s computation guide for contact mechanics, then MIT’s Robotic Manipulation notes for picking and grasping.