THE MECHANISM
Behavior cloning → shift → recovery
03Behavior cloning is ordinary supervised learning
→04Small errors change the data
→05Two failures need two fixes
FAILURE → REPAIR
Common trap
“Low validation MSE means the robot is ready.”
Better model
Validation from expert data tests prediction in familiar states. A rollout tests whether the policy can survive its own state distribution.
YOUR NEXT GATE
Feel covariate shift on a 2D toy
- 1Open the ANIMA foundations lab.
- 2Train the MSE behavior-cloning baseline.
- 3Roll it out, not just validate it.