The paper presents a novel approach to imitation learning (IL) in robotics called CasIL (Cognition-Action-based Skill Imitation Learning). CasIL introduces human cognitive priors into a dual cognition-action architecture, enabling robots to effectively learn and imitate complex skills from raw visual demonstrations. The approach aims to improve the robustness and reliability of skill imitation in long-horizon tasks such as locomotion and manipulation. The authors evaluate their method using benchmarks and tasks for quadrupedal robot locomotion, demonstrating competitive and robust skill imitation capabilities.
Publication date: 2 Oct 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2309.16299
