Hand-in-the-Loop: Improving Dexterous VLA via Seamless Interventional Correction

Zhuohang Li, Liqun Huang, Wei Xu, Zhengming Zhu, Nie Lin et al.|May 14, 2026arXiv

Key Takeaway

Seamlessly blending human intervention with robot policy execution—rather than abrupt takeovers—dramatically reduces manipulation failures in dexterous tasks and produces better-trained policies from human correction data.

Summary

This paper addresses a key problem in robotic hand control: when humans take over from an AI policy during manipulation tasks, abrupt hand configuration changes ('gesture jumps') cause failures. Hand-in-the-Loop smoothly blends human corrections with the robot's ongoing actions, reducing takeover disruptions by 99.8% and improving task success rates by 19% when used to train better policies.

agents training

Key Terms

vision-language-action-model dexterous-manipulation interactive-imitation-learning human-in-the-loop contact-rich-dynamics