ORPA framework enables real-time robot corrections without retraining
A lightweight module fixes imitation learning errors on the fly, boosting success rates on ALOHA.
A new paper presents Online Residual Policy Adaptation (ORPA), a framework that enables immediate, feedback-driven correction of robot actions without modifying the underlying policy parameters. ORPA targets imitation-learned manipulation policies like Action Chunking with Transformers (ACT), which can be sensitive to execution errors and distribution shifts. It augments a pretrained control policy with a lightweight, feedback-conditioned module that predicts residual adjustments directly in joint space, allowing runtime adaptation. Evaluated on precision-sensitive manipulation tasks with the ALOHA platform, ORPA improved success rates and recovery from small perturbations compared to baseline control policies and rule-based inverse kinematics corrections.
- ORPA adds a lightweight feedback-conditioned module to pretrained policies like ACT, predicting joint-space residuals for real-time correction.
- Evaluated on ALOHA precision manipulation tasks, it improved success rates and recovery from small perturbations versus baselines.
- No retraining or dataset aggregation is needed—adaptation happens at runtime, enabling deployment in dynamic environments.
Why It Matters
ORPA makes robot policies practical for real-world deployment by fixing errors instantly without costly retraining.