VR-DAgger: Immersive VR for Dexterous Data Collection and Uncertainty-Guided On-Policy Correction

2026-05-26

AI Summary

This work addresses key challenges in robotic imitation learning—namely, the scarcity of task-specific demonstration data, performance degradation due to distributional shift, and inefficient expert intervention—by proposing an online policy correction framework that integrates immersive virtual reality (VR) with an uncertainty-guided mechanism. The approach leverages VR to efficiently collect high-fidelity demonstrations of dexterous manipulation tasks and employs policy uncertainty estimates to identify critical failure cases, thereby directing expert feedback toward the most informative interventions. This strategy substantially improves both data collection efficiency and the utility of expert corrections, effectively mitigating performance deterioration caused by distributional shift. Empirical results demonstrate that the method achieves superior learning efficiency and final task performance on complex dexterous manipulation benchmarks.

Abstract

Learning from demonstrations is effective for robotic manipulation, but collecting sufficient task-specific data remains a major bottleneck. Under distribution shift, small errors compound, performance degrades, and expert time is often spent on redundant, low-value corrections instead of the few critical failure cases.

Problem

• robotic manipulation
• data collection
• distribution shift
• expert corrections
• learning from demonstrations

Innovation

• Immersive VR
• Dexterous Manipulation
• Uncertainty-Guided Correction
• On-Policy Learning
• Demonstration Learning