CompliantBot: Learning Whole-Body Compliance across Task Interfaces for Humanoid Control

This research presents CompliantBot, a compliance-training framework designed for humanoid robots to yield to external forces while retaining the ability to resume original tasks. The method uses a state-dependent interaction force field and complementary rewards to coordinate yielding and task resumption without requiring preconstructed compliant data or explicit external-force measurements. Evaluations show that CompliantBot yields farther under constant pulls, experiences lower spring interaction loads, achieves higher task completion rates under stronger pulls, and consumes the lowest mean joint mechanical power compared to matched baselines. Real-robot demonstrations confirm compliant responses in arm motion, stepping, and body reorientation.