Paper
Code
Video
Abstract
Legged robots are increasingly sought for tasks that require both agile locomotion and versatile manipulation, yet equipping them with dedicated arms often adds weight and complexity. In this work, we introduce a hierarchical reinforcement learning framework for pedipulation---using a quadruped's legs for manipulation tasks. Our approach comprises a high-level policy that dynamically selects which leg to use, and a low-level policy that learns precise foot control, including stable contact force when interacting with rigid objects. By combining these two levels of control, we significantly reduce execution time when tasks (e.g., pressing a button) appear on different sides, while maintaining robust balance and accuracy. Crucially, we propose a novel reward design that encourages both reaching the target point and applying the correct contact force to ensure successful and safe manipulation. Simulation results show that our hierarchical method outperforms single-limb approaches in both efficiency and fine-force control, demonstrating the feasibility of using multiple legs for rapid, precise, and stable interactions in dynamic environments.
Method
We propose a hierarchical reinforcement learning (HRL) framework, which can be divided into two parts: the high-level policy and the low-level policy. The high-level policy determines which foot—right front foot or left front foot—should be used. The low-level policy is responsible for controlling a single foot to reach the target point or exert a precise force on a specific object.
