ORPA: Real-Time Residual Policy Adaptation for Robot Manipulation

The recently introduced framework, ORPA—Online Residual Policy Adaptation—facilitates immediate adjustments to robotic manipulation actions without the need for retraining the core policy. Tailored for imitation learning strategies such as Action Chunking with Transformers (ACT), ORPA features a feedback-conditioned module that forecasts necessary residual modifications in joint space, enabling real-time behavioral changes. Conventional methods necessitate expensive dataset collection and complete policy retraining, which are impractical for real-time applications. By leveraging human feedback, ORPA refines pretrained control policies for instantaneous corrections. Tested on precision-critical tasks utilizing the ALOHA platform, results demonstrated enhanced success rates, reflecting improved robustness. The preprint arXiv:2608.17323 outlines the methodology, module design, and evaluation findings, presenting a viable solution for adapting robot policies in ever-changing environments.