The article discusses the challenges of Offline Imitation Learning (IL) in learning visuomotor skills, particularly high-precision manipulation tasks. The authors introduce the Constrained-Context Conditional Diffusion Model (C3DM), designed to balance between sample efficiency, robustness against distractors, and solving high-precision manipulation tasks with complex action spaces. C3DM includes a fixation step that helps the action denoiser to focus on task-relevant regions. The authors demonstrate that C3DM consistently achieves high success rates across a range of tasks requiring varying levels of precision and robustness to distractors.
Publication date: 3 Nov 2023
Project Page: https://sites.google.com/view/c3dm-imitation-learning
Paper: https://arxiv.org/pdf/2311.01419
