Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance
HiRe-MoMa improved TIAGo++ task performance with around 30 minutes of on-robot training and human intervention covering under 20% of training.
The framework keeps a base policy frozen and learns additive corrections to its actions. Separate correction heads control the mobile base and manipulator arm, with one-way conditioning designed to reduce interference during learning. Sparse human interventions guide exploration by adjusting the critic’s value estimates. The authors report improvements across the base policy architectures tested, though the abstract gives no numerical success rates.
ArXiv · AI/CL/LG's note
The framework keeps a base policy frozen and learns additive corrections to its actions. Separate correction heads control the mobile base and manipulator arm, with one-way conditioning designed to reduce interference during learning. Sparse human interventions guide exploration by adjusting the critic’s value estimates. The authors report improvements across the base policy architectures tested, though the abstract gives no numerical success rates.
ArXiv · AI/CL/LG's note
score 5