ResMimic

Technical Analysis Humanoid robots performing tasks like a human with ResMimic Recent advancements have made humanoid robots capable of moving and acting like humans. However, these approaches lack precision in movements. Here comes "ResMimic: From General Motion Tracking to Humanoid Whole-body Loco-Manipulation via Residual Learning", a two-stage residual-learning designed for precise and expressive humanoid control from human motion data. Motivation of the Work Recent advancements in General Motion Tracking (GMT) policies have allowed humanoids to successfully copy or reproduce diverse human movements using large-scale human-only motion datasets. However they still have some limitations like: Lack of Object Awareness: Existing GMT policies, while good at movement, are "unaware of manipulated objects". If a human motion involves holding a box, the GMT policy can mimic the body position but doesn't know how to precisely interact with the physical box itself. The Embodiment Gap: When researchers try to translate a human's movements directly to a robot (a process called retargeting), there are physical differences (the embodiment gap) between the human and the robot that cause failures. This results in undesirable errors, such as the robot's hand penetrating the object or floating away without contact. Lack of Generality: Many current humanoid loco-manipulation methods rely on highly specific designs (like dividing the task into stages or using handcrafted data pipelines),…