HiRe-MoMa
A hierarchical residual learning framework for fine-tuning real-world mobile manipulation robots. It freezes the base policy and learns only residual corrections for the mobile base and the robotic arm, combined with sparse human-guided value alignment. This design enables high sample efficiency when adapting policies to real-world mobile manipulation tasks.
Timeline (last 90 days)
HiRe-MoMa is a systematic framework for fine-tuning real-world mobile manipulation (MoMa) systems, adopting a hierarchical residual architecture.
The core design of HiRe-MoMa is to freeze an existing base policy and learn only additive correction residuals on top of its output, rather than training a mobile manipulation policy from scratch.
HiRe-MoMa splits the residual into two heads, the mobile base and the robotic arm, and adopts unidirectional conditioning, i.e., the base first determines the overall pose, and the arm then completes fine manipulation on that basis.
HiRe-MoMa adopts an expert-guided value alignment method, using sparse human interventions to reshape the critic's value landscape.
In HiRe-MoMa, the training coverage of human interventions is less than 20%.
HiRe-MoMa conducted extensive real-world experiments on the TIAGo++ platform, and regardless of the architecture used for the base policy, it can stably improve task performance after about 30 minutes of online interaction.
The authors state that, to their knowledge, HiRe-MoMa is among the first frameworks capable of achieving efficient fine-tuning on mobile manipulation robots in the real world.
The final performance ceiling of the HiRe-MoMa framework is limited by the quality of the frozen base policy itself.
The effectiveness of HiRe-MoMa's sparse interventions depends on the timing and quality of the interventions.
The HiRe-MoMa project website is https://hire-moma.github.io/.
All Facts (10)
HiRe-MoMa is a systematic framework for fine-tuning real-world mobile manipulation (MoMa) systems, adopting a hierarchical residual architecture.
50%UnverifiedThe core design of HiRe-MoMa is to freeze an existing base policy and learn only additive correction residuals on top of its output, rather than training a mobile manipulation policy from scratch.
50%UnverifiedHiRe-MoMa splits the residual into two heads, the mobile base and the robotic arm, and adopts unidirectional conditioning, i.e., the base first determines the overall pose, and the arm then completes fine manipulation on that basis.
50%UnverifiedHiRe-MoMa adopts an expert-guided value alignment method, using sparse human interventions to reshape the critic's value landscape.
50%UnverifiedIn HiRe-MoMa, the training coverage of human interventions is less than 20%.
50%UnverifiedHiRe-MoMa conducted extensive real-world experiments on the TIAGo++ platform, and regardless of the architecture used for the base policy, it can stably improve task performance after about 30 minutes of online interaction.
50%UnverifiedThe authors state that, to their knowledge, HiRe-MoMa is among the first frameworks capable of achieving efficient fine-tuning on mobile manipulation robots in the real world.
50%UnverifiedThe final performance ceiling of the HiRe-MoMa framework is limited by the quality of the frozen base policy itself.
50%UnverifiedThe effectiveness of HiRe-MoMa's sparse interventions depends on the timing and quality of the interventions.
50%UnverifiedThe HiRe-MoMa project website is https://hire-moma.github.io/.
50%