Jeonghyeon Kim, Shinsuk Park (corresponding)

Abstract

A humanoid’s whole-body dynamics changes continuously with upper-body motion, payload, and external interaction, but proprioception alone doesn’t explicitly represent the resulting balance demand — most policies are left to infer it implicitly. This work derives a Balance Demand Descriptor (BDD) from centroidal dynamics that makes that demand explicit, and integrates it into a reinforcement-learning locomotion policy so the policy can distinguish disturbances it can absorb within its current stance from those that require a corrective step.

Method

• Estimated the balance demand induced by upper-body motion, payload, and external interaction from centroidal momentum, formalized as the Balance Demand Descriptor (BDD)
• Integrated BDD into an RL locomotion policy’s observation so it can explicitly distinguish demands manageable within the current stance from those requiring corrective stepping
• Designed reward shaping that induces an appropriate recovery strategy based on the direction and magnitude of a disturbance

Status

Validating the physical plausibility of BDD and evaluating adaptation performance across locomotion and loco-manipulation scenarios; manuscript in preparation.

Skills

Python Reinforcement Learning Isaac Gym Centroidal Dynamics