Letian Fu

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2023-107

May 11, 2023

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-107.pdf

Grasping and insertion represent two fundamental skills for robots, garnering significant interest within the robotics community due to their widespread applications in fields such as manufacturing, logistics, maintenance, and repair. Although numerous studies have demonstrated success in both tasks, several challenges persist. For instance, general-purpose, learning-based grasping systems often struggle to identify optimal grasps for novel, out-ofdistribution industrial components, necessitating manual predefinition by humans. Likewise, many learning-based insertion algorithms require extensive demonstrations from human teleoperators and assume fixed grasp poses with minimal rotation, limiting their adaptability. Addressing these limitations, we study the problem of grasp identification and industrial insertion through two different learning-based approaches that can directly operate the physical robot with minimal human interventions, both of which achieved better performance than baselines in their respective tasks.

Advisors: Ken Goldberg

---

BibTeX citation:

@mastersthesis{Fu:EECS-2023-107,
    Author= {Fu, Letian},
    Title= {Autonomous Learning for Industrial Manipulation: Enhancing Grasping and Insertion Tasks through Scalable Data Collection},
    School= {EECS Department, University of California, Berkeley},
    Year= {2023},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-107.html},
    Number= {UCB/EECS-2023-107},
    Abstract= {Grasping and insertion represent two fundamental skills for robots, garnering significant interest within the robotics community due to their widespread applications in fields such as manufacturing, logistics, maintenance, and repair. Although numerous studies have demonstrated success in both tasks, several challenges persist. For instance, general-purpose, learning-based grasping systems often struggle to identify optimal grasps for novel, out-ofdistribution industrial components, necessitating manual predefinition by humans. Likewise, many learning-based insertion algorithms require extensive demonstrations from human teleoperators and assume fixed grasp poses with minimal rotation, limiting their adaptability. Addressing these limitations, we study the problem of grasp identification and industrial insertion through two different learning-based approaches that can directly operate the physical robot with minimal human interventions, both of which achieved better performance than baselines in their respective tasks.},
}

EndNote citation:

%0 Thesis
%A Fu, Letian 
%T Autonomous Learning for Industrial Manipulation: Enhancing Grasping and Insertion Tasks through Scalable Data Collection
%I EECS Department, University of California, Berkeley
%D 2023
%8 May 11
%@ UCB/EECS-2023-107
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-107.html
%F Fu:EECS-2023-107