Two AI agents, same world and instruction, different training data.

A hobby and research project exploring how imitation learning is best handled with a limited supervision budget.

I started Grounded Recovery after my master’s thesis to explore imitation learning and its use in robotics. I was particularly interested in a practical question: when expert labels are limited, which situations are most worth learning from?

Alongside the experiments, I built interactive visualizations and small lessons to work through the fundamentals myself. It’s something I enjoy, and often how abstract ideas start to make sense to me.

This website was created for curious minds engaging with imitation learning: a short, visual introduction to imitation learning and DAgger-style approaches. I hope it offers a useful starting point for others coming to the subject, too.

Two ways to spend the same labels

Expert demonstrations cover a narrow band of states. A learner trained on them knows what to do inside it.
Extra demonstrations spend a label budget inside that band, on the kind of states the learner has already seen.
Deployment: one wrong action can take the learner off the band, into states no demonstration covers. There, small errors compound.
Corrections spend the same budget on the states the learner reaches after such a mistake, each labelled with the expert’s way back.

Which of the two helps more is what the study measures.

How are limited expert labels best spent on imitation learning for robots?

An agent can learn from expert demonstrations or from advice after a mistake. With the same number of labels, which helps more? This project compares both in a small simulated maze.