Safe Whole-Body Loco-Manipulation via Combined Model and Learning-based Control

A research team from MIT and UC Berkeley has published a paper, 'Safe Whole-Body Loco-Manipulation via Combined Model and Learning-based Control,' proposing a breakthrough controller for legged robots. The work, accepted to the prestigious IEEE International Conference on Robotics and Automation (ICRA) 2026, tackles the critical challenge of enabling robots to walk and manipulate objects simultaneously while remaining safe and compliant during physical contact with humans or the environment. This moves beyond the limitations of stationary robotic arms, aiming for true mobile manipulation.

The controller's key innovation is its hybrid architecture. It uses a model-based admittance controller for the 6-DoF manipulator arm, which maps external forces (wrenches) into desired motion, creating compliant, human-like interaction. This is paired with a learned Reinforcement Learning (RL) policy to handle the complex dynamics of legged locomotion. A Reference Governor (RG) provides formal safety guarantees, and a neural network-enhanced Kalman filter improves state estimation. The team successfully validated the system on a Unitree Go2 quadruped equipped with a custom arm and force/torque sensor, demonstrating reliable 6-DoF force response in dynamic settings. This paves the way for robots that can safely assist in warehouses, construction sites, or homes by physically interacting with their surroundings while on the move.