ResMerge: Residual Policy Learning and Merging for Continual Adaptation of Pre-trained Robot Policies
ResMerge: Residual Policy Learning and Merging for Continual Adaptation of Pre-trained Robot Policies
This research introduces ResMerge, a framework designed for the continual adaptation of pre-trained robot policies. It employs a two-stage approach involving residual reinforcement learning to acquire task-specific corrections without needing new expert demonstrations, followed by parameter-space merging to consolidate compatible residuals and reduce storage requirements. The method includes a metric called Residual-Field Consistency (RFC) to assess compatibility in action-field space. Experiments conducted on the LIBERO benchmark across two different backbones and on a real-world Franka robot demonstrate that ResMerge achieves improved continual learning performance compared to strong Parameter-Efficient Continual Fine-Tuning (PECFT) baselines, while also reducing the residual-parameter footprint through selective merging.
The paper proposes ResMerge, a two-stage framework combining residual reinforcement learning and parameter-space merging to enable continual adaptation of pre-trained robot policies without requiring additional expert demonstrations for each new task.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- code: not_reported
- weights: not_reported
- dataset: not_reported
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
erpretable metric that assesses compatibility in residual action-field space by comparing correction direction and magnitude. Experiments on LIBERO across two backbones and on a real-world Franka robot show improved continual learning performance over strong PECFT baselines, while selective merging reduces the residual-parameter footprint. Project website: https://resmerge.github.io/.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org