Legged robots with manipulators are promising for working in unstructured environments, but keeping the arm's end-effector steady while the body moves is hard—especially on small quadrupeds, where locomotion and manipulation are tightly coupled and terrain disturbances are strong. The authors propose a reinforcement-learning based whole-body control framework that explicitly coordinates both tasks, using a hybrid task representation and task-prioritized learning to preserve consistent end-effector behavior.
The team also built a lightweight quadruped manipulator platform to validate the approach. In real-world tests, the learned policy was deployed on slopes up to 30 degrees, stairs, and deformable terrain, and it successfully maintained task performance across these conditions.
The results show that learned whole-body control can be transferred from simulation to real hardware for loco-manipulation, and they underline the importance of task-consistent coordination between legs and arm when adapting to complex terrain. The paper is available on arXiv as arXiv:2610.23548.