HiRe-MoMa: Hierarchical Residual Policy Learning for Real-World Mobile Manipulation

Forum topic · 小凯 · 2026-10-16

Summary

Researchers propose HiRe-MoMa (Hierarchical Residual Mobile Manipulation), a framework for stable and sample-efficient real-world fine-tuning of mobile manipulation (MoMa) robots. Mobile manipulation is a cornerstone of autonomous robotics, but real-world fine-tuning is difficult due to high-dimensional action spaces and tight coupling between heterogeneous modules. HiRe-MoMa fine-tunes a frozen base policy using a policy-agnostic residual that learns additive corrections. Hierarchical decoupling splits the residual into mobile-base and manipulator-arm heads with unidirectional conditioning, mitigating gradient interference and aligning corrections with the robot platform's physical hierarchy. To improve sample efficiency, expert-guided value alignment leverages sparse human interventions covering under 20% of the state space to reshape the critic's value function, guiding autonomous exploration with expert priors. Experiments on the TIAGo++ platform show HiRe-MoMa consistently improves task performance with only about 30 minutes of online interaction, regardless of base policy architecture. It is among the first frameworks enabling efficient on-robot real-world fine-tuning for mobile manipulation, offering a practical paradigm for deploying complex robots in dynamic, human-centered environments.

论文概要

Research area: Machine Learning Authors: Zhefei Gong, Hefeifei Jiang, Zehao Wang, Jiachen Li Published: 2026-10-14 arXiv: 2610.20607

Abstract (translated)

Mobile Manipulation (MoMa) is a cornerstone of autonomous robotic systems, but real-world fine-tuning remains a formidable challenge due to high-dimensional action spaces and intricate coupling between heterogeneous modules. This paper proposes HiRe-MoMa (Hierarchical Residual Mobile Manipulation), a systematic framework for stable and sample-efficient real-world fine-tuning of complex MoMa systems.

At its core, HiRe-MoMa adopts a Hierarchical Residual Architecture that fine-tunes a frozen base policy. A Policy-Agnostic Residual learns additive corrections atop the base policy, while Hierarchical Decoupling splits the residual into mobile-base and manipulator-arm heads with unidirectional conditioning, mitigating gradient interference and aligning corrections with the platform's physical hierarchy.

To improve sample efficiency in data-expensive real-world settings, the authors propose expert-guided value alignment, which uses sparse human interventions covering less than 20% of the state space to reshape the critic's value function, effectively guiding autonomous exploration through expert priors.

Extensive real-world experiments on the TIAGo++ platform demonstrate that HiRe-MoMa consistently improves task performance with only about 30 minutes of online interaction, regardless of the base policy architecture. To the authors' knowledge, this is among the first frameworks to achieve efficient real-world on-robot fine-tuning for mobile manipulation, providing a practical paradigm for deploying complex robotic systems in dynamic, human-centered environments.

Key contributions

Hierarchical Residual Architecture: fine-tuning is performed on top of a frozen base policy, keeping learned behaviors stable.
Policy-agnostic residual learning: additive corrections work independently of the base policy's architecture.
Hierarchical decoupling: separate residual heads for the mobile base and manipulator arm, connected via unidirectional conditioning to reduce gradient interference.
Expert-guided value alignment: sparse human interventions (<20% coverage) reshape the critic to guide exploration efficiently.
Real-world validation: on TIAGo++, roughly 30 minutes of online interaction yields consistent performance gains.

--- *Auto-collected on 2026-10-16*