Robot Imitation Learning

Robot imitation learning studies how robots can acquire manipulation, navigation, locomotion, and whole-body skills from demonstrations rather than hand-coded rules or task rewards. Common approaches include behavior cloning with transformers or diffusion policies, interactive methods such as DAgger, inverse reinforcement learning, and hierarchical skill learning; demonstrations may come from teleoperation, human video, simulation, or robot-free interfaces. Research also uses visual, tactile, force, and proprioceptive feedback, with motion retargeting and data augmentation helping transfer skills across robots and settings. A central challenge is making policies reliable beyond their demonstrations: distribution shift, imperfect data, compounding errors, and sim-to-real gaps remain, alongside the need for efficient data collection and safe recovery.