Robot Learning 2026 – Lecture 3: Imitation Learning | ETH Zürich

Imitation learning enables robots to acquire skills by mimicking expert trajectories, yet naive implementations like behavior cloning often suffer from distributional shift and compounding errors. When a policy deviates from the expert's path, it encounters unfamiliar states, leading to quadratic error growth. Mitigating these issues requires techniques like dataset aggregation (Dagger) or online human interventions to guide recovery. Furthermore, because human demonstrations are often non-Markovian and multi-modal—where multiple valid actions exist for a single state—effective policies must incorporate history or utilize advanced generative frameworks like diffusion and latent variable models. By conditioning on goal states rather than fixed tasks, robots can learn more robust, task-agnostic control from unstructured play data, effectively bridging the gap between simple imitation and complex, real-world robotic performance.