FlowCorrect: Efficient Interactive Correction of Generative Flow Policies for Robotic Manipulation
FlowCorrect uses VR interface for human corrections, boosting robotic manipulation success to 80%.
Key Findings
Methodology
FlowCorrect is a modular interactive imitation learning approach that adapts generative flow policies using sparse human corrections. During execution, humans provide brief corrective pose nudges via a lightweight VR interface. FlowCorrect utilizes these sparse corrections to locally adapt the policy, improving actions without retraining the backbone while preserving model performance on previously learned scenarios.
Key Results
- FlowCorrect achieved an 80% success rate across four tabletop tasks, notably improving previously failed cases while maintaining performance on solved scenarios.
- Compared to full-policy retraining, FlowCorrect excels in efficiency, enabling rapid deployment-time recovery of failed cases.
- With human corrections, FlowCorrect enables fast, sample-efficient, incremental human-in-the-loop corrections under low correction budgets.
Significance
The significance of FlowCorrect lies in addressing execution failures caused by deployment-time distribution shifts in generative manipulation policies. By allowing human corrections, FlowCorrect can quickly adapt to new situations without retraining, maintaining the stability and performance of the base policy. This approach opens new possibilities for real-time adjustment of robotic manipulation policies.
Technical Contribution
FlowCorrect introduces a lightweight correction module that allows localized adjustments without affecting the original policy. Compared to existing state-of-the-art methods, FlowCorrect provides new theoretical guarantees and engineering possibilities, especially in handling distribution shifts and low correction budgets.
Novelty
FlowCorrect is the first to achieve deployment-time adaptation of generative flow policies through sparse human corrections, offering a more natural and intuitive interaction compared to existing absolute correction methods.
Limitations
- FlowCorrect may struggle in extreme distribution shift scenarios due to the need for sufficiently representative correction data.
- The method relies on human correction input, which may increase cognitive load.
- Corrections may require higher precision in complex tasks.
Future Work
Future work could explore reducing human correction burden, developing smarter automatic correction mechanisms, and extending to more complex robotic manipulation tasks.
AI Executive Summary
FlowCorrect is an innovative interactive imitation learning method designed to address distribution shift issues in robotic manipulation policies during deployment. Traditional generative policies often fail when test-time states differ from those seen during training. FlowCorrect allows humans to provide brief corrective pose nudges via a lightweight VR interface, enabling local policy adaptation without retraining.
In experiments, FlowCorrect achieved an 80% success rate across four tabletop tasks, notably improving previously failed cases while maintaining performance on solved scenarios. Compared to full-policy retraining, FlowCorrect excels in efficiency, enabling rapid deployment-time recovery of failed cases. With human corrections, FlowCorrect enables fast, sample-efficient, incremental human-in-the-loop corrections under low correction budgets.
The significance of FlowCorrect lies in addressing execution failures caused by deployment-time distribution shifts in generative manipulation policies. By allowing human corrections, FlowCorrect can quickly adapt to new situations without retraining, maintaining the stability and performance of the base policy. This approach opens new possibilities for real-time adjustment of robotic manipulation policies.
Deep Dive
Abstract
Generative manipulation policies can fail catastrophically under deployment-time distribution shift, yet many failures are near-misses: the robot reaches almost-correct poses and would succeed with a small corrective motion. We propose FlowCorrect, a modular interactive imitation learning approach that enables deployment-time adaptation of flow-matching manipulation policies from sparse, relative human corrections without retraining. During execution, a human provides brief corrective pose nudges via a lightweight VR interface. FlowCorrect uses these sparse corrections to locally adapt the policy, improving actions without retraining the backbone while preserving the model performance on previously learned scenarios. We evaluate on a real-world robot across four tabletop tasks: pick-and-place, pouring, cup uprighting, and insertion. With a low correction budget, FlowCorrect achieves an 80% success rate on previously failed cases while preserving performance on previously solved scenarios. The results clearly demonstrate that FlowCorrect learns from very few demonstrations and enables fast, sample-efficient, incremental, human-in-the-loop corrections of generative visuomotor policies at deployment time in real-world robotics.