SEEC: Stable End-Effector Control

Humanoid loco-manipulation requires keeping an end-effector stable while the whole body moves. SEEC pairs a model-based controller with residual reinforcement learning so the policy only has to learn what the model gets wrong, rather than re-learning the full task from scratch.

Overview

The model-based component provides a stabilizing prior for end-effector tracking, and a learned residual corrects for unmodeled dynamics during locomotion. The result is a controller that handles loco-manipulation tasks more reliably than either component alone, while remaining sample-efficient enough to train.

Publication

Jaehwi Jang, Zhuoheng Wang, Ziyi Zhou, Feiyang Wu, Ye Zhao
ICRA 2026

Links