Errors and recovery
A mistake can take the learner somewhere its expert demonstrations rarely visit. Which examples would help it recover?
After one changed action
We replace one chosen action with another, such as a turn instead of a forward step. We then measure success and distance from the expert’s route.
Testing a change the agent has not trained on
During recovery training, left becomes right, right becomes forward, and forward becomes left. The main test reverses this mapping and uses different times. We call these forced changes corruptions.
“Unseen” means the new mapping and timing were absent from training. This tests one new error pattern.
Where should we ask for advice?
Both methods start with the same agent and get more labels. Extra demonstrations follow expert routes. Recovery labels advise the learner for a few steps after a forced change.
DAgger repeats a simple loop: run the learner, ask the expert what to do in the states it visits, add the labels, and retrain. Here, we focus that advice on steps after an action change.
Equal label counts are only part of a fair comparison
Labels can be reused during training. Here, each update contains more new labels for one method than the other. The study fixes this by matching labelled targets per update.
What happens in the small experiment?
Move through the episode below. The amber frame marks the changed action. Watch what the agent sees and does next.
Why the study needs more runs
The three recovery-minus-demonstration differences are . Neither method wins consistently. The study repeats this comparison with more runs and equal training exposure.