Bimanual Dexterity: 3D Object Reconstruction and Cross-Embodiment Learning for Generalizable Manipulation

Zehan Ma

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-98
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-98.pdf

Bimanual robots are capable of performing complex tasks that require coordination and dexterity, such as folding, handovers, and assembly. In addition to their utility in task execution, bimanual platforms also offer unique advantages for generating data to support scalable perception and policy learning. This thesis explores how dual-arm robots can be leveraged to support generalizable manipulation through two complementary systems. To address the challenge of creating complete 3D object models suitable for downstream tasks, we present a method that uses coordinated in-hand scanning and regrasping to produce high-fidelity, occlusion-free 3D Gaussian Splat reconstructions from a fixed camera. Meanwhile, to overcome the scarcity of bimanual training data, we introduce a cross-embodiment learning framework that trains dual-arm policies from single-arm teleoperation, using role alternation and vision-based synthesis to generate full bimanual demonstrations. Together, these systems demonstrate how bimanual robots can facilitate scalable data generation in both perception and policy learning, reducing reliance on specialized hardware and manual supervision while improving generalization across tasks and embodiments.

Advisor: Ken Goldberg

\"Edit"; ?>

BibTeX citation:

@mastersthesis{Ma:EECS-2025-98,
    Author = {Ma, Zehan},
    Title = {Bimanual Dexterity: 3D Object Reconstruction and Cross-Embodiment Learning for Generalizable Manipulation},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-98.html},
    Number = {UCB/EECS-2025-98},
    Abstract = {Bimanual robots are capable of performing complex tasks that require coordination and dexterity, such as folding, handovers, and assembly. In addition to their utility in task execution, bimanual platforms also offer unique advantages for generating data to support scalable perception and policy learning. This thesis explores how dual-arm robots can be leveraged to support generalizable manipulation through two complementary systems. To address the challenge of creating complete 3D object models suitable for downstream tasks, we present a method that uses coordinated in-hand scanning and regrasping to produce high-fidelity, occlusion-free 3D Gaussian Splat reconstructions from a fixed camera. Meanwhile, to overcome the scarcity of bimanual training data, we introduce a cross-embodiment learning framework that trains dual-arm policies from single-arm teleoperation, using role alternation and vision-based synthesis to generate full bimanual demonstrations. Together, these systems demonstrate how bimanual robots can facilitate scalable data generation in both perception and policy learning, reducing reliance on specialized hardware and manual supervision while improving generalization across tasks and embodiments.}
}

EndNote citation:

%0 Thesis
%A Ma, Zehan
%T Bimanual Dexterity: 3D Object Reconstruction and Cross-Embodiment Learning for Generalizable Manipulation
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-98
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-98.html
%F Ma:EECS-2025-98