Tactile-augmented agent for slip-related dexterous manipulation
Author(s)
Chen, Tianyi
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Contact-rich robotic grasping is crucial for manipulating objects in various industrial and household applications. In this thesis, we focus on a fluid collection task involving a multi-fingered robotic hand with 24 degrees of freedom and investigate the effects of incorporating tactile information into both sensing and the reward function. To this end, we designed and trained multiple reinforcement learning agents, each differing from the others in terms of how tactile information is incorporated into the sensing or reward signals. Specifically, we consider four agents: i) Baseline: no tactile information; ii) Tactile-Aware (TA): the ability to sense tactile information; iii) Tactile-Reward (TR): tactile information influencing rewards; iv) TS + TR: tactile information incorporating into both sensing and reward signals. Our results demonstrate that the ability to sense tactile information (i.e., the TS agent) can significantly improve the learning speed and final reward of the agents compared to the baseline agent. However, we found that adding tactile information to the reward function can decrease sample efficiency and lead to slower learning. Further, our qualitative analysis shows that the presence of tactile information and reward can affect the learned grasping gesture. Specifically, we find that leveraging tactile information for reward signals leads to more natural and ”human-like” grasps compared to agents that do not utilize shaped rewards that incorporate tactile information. Our study highlights the importance of considering tactile feedback in robotic grasping tasks and provides insights into the effects of different forms of tactile information and reward on learning performance.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis