Whole-Body Dynamic Throwing with Legged Manipulators

2024-10-08

arXiv.org

AI Summary

This study addresses the fundamental challenge in legged manipulation robots—namely, the decoupling of locomotion and manipulation capabilities, and the trade-off between throwing accuracy and dynamic stability—during dynamic throwing tasks. We propose the first end-to-end throwing control framework integrating full-body dynamics, encompassing leg-driven propulsion, momentum transfer, and reactive balance regulation. Methodologically, we develop a PPO-based policy trained on full-state observations, employing a sparse-reward environment and adaptive curriculum learning to achieve high-precision, stable 3D target-reaching throws. The framework demonstrates successful sim-to-real transfer from PyBullet/MuJoCo simulations to a physical humanoid robot. Experiments show substantial improvements: 42% increase in throwing distance, sub-0.15 m positional error, and zero tip-over incidents—while maintaining robustness across diverse targets and platforms. This work establishes a new paradigm for embodied agents performing complex, coordinated manipulation-locomotion tasks.

Technology Category

• Intelligent Robots: Manipulation
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Machine Learning: Imitation Learning & Inverse Reinforcement Learning

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Agentic search

Abstract

Throwing with a legged robot involves precise coordination of object manipulation and locomotion - crucial for advanced real-world interactions. Most research focuses on either manipulation or locomotion, with minimal exploration of tasks requiring both. This work investigates leveraging all available motors (full-body) over arm-only throwing in legged manipulators. We frame the task as a deep reinforcement learning (RL) objective, optimising throwing accuracy towards any user-commanded target destination and the robot's stability. Evaluations on a humanoid and an armed quadruped in simulation show that full-body throwing improves range, accuracy, and stability by exploiting body momentum, counter-balancing, and full-body dynamics. We introduce an optimised adaptive curriculum to balance throwing accuracy and stability, along with a tailored RL environment setup for efficient learning in sparse-reward conditions. Unlike prior work, our approach generalises to targets in 3D space. We transfer our learned controllers from simulation to a real humanoid platform.

Problem

• Coordination of manipulation and locomotion in legged robots for throwing tasks
• Full-body dynamic throwing vs. arm-only throwing in legged manipulators
• Generalization to 3D target destinations with real-world transfer

Innovation

• Full-body motor coordination for dynamic throwing
• Deep reinforcement learning optimizes accuracy and stability
• Adaptive curriculum balances learning in sparse-reward conditions