CHG-DAgger: Interactive Imitation Learning with Human-Policy Cooperative Control

Taro Takahashi, Yutaro Ishida, Takayuki Kanai, Naveen Kuppuswamy

Keywords: Interactive Imitation Learning, Collaborative Control, Retraining, Visuomotor Control Policy

TL;DR: Proposed method enables humans and learned policy to collaborate in correcting task failures without switching control, reducing intervention time.

Abstract: This paper presents a novel approach to improve the usability of Interactive Imitation Learning (IIL) for robot motion generation. The proposed framework, Cooperative-HG-DAgger (CHD-DAgger), allows the expert human and learned policy to collaborate in continuing the task upon task failure without switching control between the policy and human. As a result, human intervention time is reduced because the human can correct the motion while being guided by the policy, and they can understand when corrections are no longer needed through physical interaction. To achieve cooperative control, we adopted multilateral control, an extension of bilateral control, designed to avoid instability even with low-cost hardware and long reference trajectory update cycles. The proposed method achieved high success rates through retraining using the recovery motion data. Additionally, it was shown that intervention time can be reduced in minor adjustment where human operation is close to the policy, when prior knowledge of the policy is limited. Our results indicate that the proposed method offers a more intuitive and efficient way of handling task failures, paving the way for continuous learning and robust robot autonomy.

Submission Number: 15