Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance

TL;DR

This paper proposes HiRe-MoMa, a framework for real-world fine-tuning of mobile manipulation systems. It achieves stable and sample-efficient fine-tuning with sparse human guidance. HiRe-MoMa enables efficient on-robot fine-tuning with only 30 minutes of online interaction.

Why It Matters

HiRe-MoMa has significant practical implications for deploying complex robotic systems in dynamic environments. It enables efficient fine-tuning of mobile manipulation systems, which is crucial for real-world applications. This framework can be used in various human-centric environments, such as warehouses or homes.

Key Findings

• Hierarchical Residual Architecture improves fine-tuning efficiency
• Expert-Guided Value Alignment enhances sample efficiency
• HiRe-MoMa achieves consistent task performance with limited online interaction