Simeon Nedelchev
Training uses demonstrations.
Execution uses the student's own observations and actions.
Complete episodes retain reset conditions and termination reasons.
Training and evaluation use different episodes.
Normalization uses training data only.
A small action error changes the next state.
The student can enter regions absent from demonstrations.
A linear student can recover a linear expert from sufficiently rich noiseless data.
This validates the pipeline before adding complex models.
Action MSE measures prediction on a dataset.
Rollout cost measures closed-loop consequences.
Both belong in the report.
Fit a small student, then test broader initial states.
Which failures appear only in rollouts?
Reading: MIT Manipulation notes