Research Analyzer
← Back ICRA 2026

RAMBO: RL-Augmented Model-Based Whole-Body Control for Loco-Manipulation

Jin Cheng, Dongho Kang, Gabriele Fadini, Guanya Shi, Stelian Coros

PDF

AI summary

Key figure (auto-extracted from paper)
RAMBO combines model-based precision with RL robustness to enable precise and stable whole-body loco-manipulation on legged robots.
loco-manipulation whole-body control reinforcement learning quadruped robot hybrid control

Problem

Legged robots struggle to balance the need for precise end-effector force control, typically provided by model-based methods, with the robustness to unmodeled dynamics offered by learning-based methods.

Approach

A hybrid framework that uses a model-based whole-body controller to generate feedforward torques via quadratic programming, augmented by an RL policy that provides corrective feedback for base acceleration and joint positions.

Key results

  • Successful real-world tasks including pushing a shopping cart, balancing a plate, and holding soft objects
  • Demonstrated capability in both quadrupedal and bipedal walking modes
  • Precise end-effector tracking while maintaining robust dynamic locomotion
  • Validated on the Unitree Go2 quadruped platform

Why it matters

This framework enables legged robots to perform complex physical interactions with their environment without sacrificing stability or precision.

Abstract

Loco-manipulation, physical interaction of vari- ous objects that is concurrently coordinated with locomotion, remains a major challenge for legged robots due to the need for both precise end-effector control and robustness to unmodeled dynamics. While model-based controllers provide precise planning via online optimization, they are limited by model inaccuracies. In contrast, learning-based methods offer robustness, but they struggle with precise modulation of inter- action forces. We introduce RAMBO, a hybrid framework that integrates model-based whole-body control within a feedback policy trained with reinforcement learning. The model-based module generates feedforward torques by solving a quadratic program, while the policy provides feedback corrective terms to enhance robustness. We validate our framework on a quadruped robot across a diverse set of real-world loco- manipulation tasks, such as pushing a shopping cart, balancing a plate, and holding soft objects, in both quadrupedal and bipedal walking. Our experiments demonstrate that RAMBO enables precise manipulation capabilities while achieving robust and dynamic locomotion.

Index terms

Legged Robots Reinforcement Learning Mobile Manipulation

Related papers