Learning Diverse Bimanual Dexterous Manipulation Skills from Human Demonstrations

Bohan Zhou,Haoqi Yuan,Yuhui Fu,Zongqing Lu

2024-10-03

Abstract:Bimanual dexterous manipulation is a critical yet underexplored area in robotics. Its high-dimensional action space and inherent task complexity present significant challenges for policy learning, and the limited task diversity in existing benchmarks hinders general-purpose skill development. Existing approaches largely depend on reinforcement learning, often constrained by intricately designed reward functions tailored to a narrow set of tasks. In this work, we present a novel approach for efficiently learning diverse bimanual dexterous skills from abundant human demonstrations. Specifically, we introduce BiDexHD, a framework that unifies task construction from existing bimanual datasets and employs teacher-student policy learning to address all tasks. The teacher learns state-based policies using a general two-stage reward function across tasks with shared behaviors, while the student distills the learned multi-task policies into a vision-based policy. With BiDexHD, scalable learning of numerous bimanual dexterous skills from auto-constructed tasks becomes feasible, offering promising advances toward universal bimanual dexterous manipulation. Our empirical evaluation on the TACO dataset, spanning 141 tasks across six categories, demonstrates a task fulfillment rate of 74.59% on trained tasks and 51.07% on unseen tasks, showcasing the effectiveness and competitive zero-shot generalization capabilities of BiDexHD. For videos and more information, visit our project page <a class="link-external link-https" href="https://sites.google.com/view/bidexhd" rel="external noopener nofollow">this https URL</a>.

Robotics,Machine Learning

What problem does this paper attempt to address?

The problem this paper attempts to address is the challenge of learning bimanual dexterous manipulation skills, which is a critical but under-researched area in robotics. Specifically, the paper focuses on how to efficiently learn diverse bimanual dexterous manipulation skills from human demonstrations. Existing methods mainly rely on reinforcement learning, but these methods are often limited by reward functions meticulously designed for narrow tasks, lacking scalability and generalization to a wide range of tasks. Therefore, the paper proposes a new framework, BiDexHD, which aims to overcome the limitations of existing methods by learning diverse bimanual dexterous manipulation skills from a large number of human demonstrations through a unified task construction approach and a teacher-student policy learning mechanism.

The main contributions of the paper include:
1. Formalizing the problem of learning bimanual dexterous skills from human demonstrations as an initial attempt towards general bimanual skills.
2. Proposing BiDexHD, a unified and scalable reinforcement learning framework for learning diverse bimanual dexterous manipulation skills from human demonstrations, enhancing the robot's ability to perform bimanual collaborative tasks.
3. Evaluating BiDexHD on the TACO dataset, covering 141 automatically constructed tasks in 6 categories, demonstrating BiDexHD's superior performance on training tasks and competitive zero-shot generalization ability to unseen tasks.