← Back ICRA 2026

A Modular Residual Learning Framework to Enhance Model-Based Approach for Robust Locomotion

Min-Gyu Kim, Dongyun Kang, Hajun Kim, Hae-Won Park

PDF

AI summary

Key figure (auto-extracted from paper)

A modular hybrid controller using RL for foothold adaptation and SL for dynamics correction significantly improves robustness and learning efficiency over pure model-based or end-to-end learning baselines.

Legged robots Model predictive control Residual learning Reinforcement learning Supervised learning Robust locomotion

Problem

Model-based locomotion controllers degrade under real-world uncertainties like payload changes and terrain variations due to necessary dynamic simplifications, while end-to-end learning methods suffer from poor sample efficiency and limited robustness.

Approach

The framework injects lightweight, decoupled learning modules into a nominal model-predictive controller: an RL network adaptively corrects foothold patterns, and an SL network predicts residual dynamics from proprioceptive sensors.

Key results

Robust task execution under heavy disturbances and OOD conditions
Reduced sensitivity to nominal controller parameter tuning
Improved learning efficiency and faster convergence compared to baselines
Successful real-world quadrupedal balance and velocity tracking

Why it matters

Provides a practical, computationally efficient pathway for deploying robust legged robots in unstructured, real-world environments without sacrificing the safety guarantees of model-based control.

Abstract

This letter presents a novel approach that combines the advantages of both model-based and learning-based frame- works to achieve robust locomotion. The residual modules are integrated with each corresponding part of the model-based frame- work, a footstep planner and dynamic model designed using heuris- tics, to complement performance degradation caused by a model mismatch.Byutilizingamodularstructureandselectingtheappro- priate learning-based method for each residual module, our frame- workdemonstratesimprovedcontrolperformanceinenvironments with high uncertainty, while also achieving higher learning effi- ciency compared to baseline methods. Moreover, we observed that our proposed methodology not only enhances control performance but also provides additional benefits, such as making nominal controllers more robust to parameter tuning. To investigate the feasibility of our framework, we demonstrated residual modules combined with model predictive control in a real quadrupedal robot. Despite uncertainties beyond the simulation, the robot suc- cessfully maintains balance and tracks the commanded velocity.

Index terms

Legged Robots Machine Learning for Robot Control Optimization and Optimal Control