Learning low-level robot manipulations skills from human demonstration videos - Ref. 99/26/IR

This project addresses video understanding for robotic applications, focusing on learning action-level representations from unstructured human activity videos to instruct humanoid robots in long-horizon manipulation tasks. It will develop hierarchical, affordance-aware, andgoal-conditioned representations of human behavior, potentially enriched with textual descriptions, whose outputs will serve as inputs to motor policies based on reinforcement and imitation learning. Experience in large-scale video understanding for human behavior analysis isrequired.