Researchers propose a method called Online Residual Policy Adaptation (ORPA) for improving the performance of robotic manipulation policies in real-time. ORPA adds a lightweight module to a pre-trained policy that predicts adjustments to actions based on feedback from humans. This allows the system to adapt its behavior immediately without retraining the entire policy. The authors tested ORPA on precision-sensitive tasks and found it improved success rates and recovery from small errors compared to baseline methods. --- Why it matters: This matters because robotic manipulation policies are often brittle and require significant computational resources to correct mistakes, making real-time deployment challenging. ORPA's ability to adapt in real-time could enable more robust and efficient control of robots in various applications. Source: https://arxiv.org/abs/2608.17323