Robot learning of manipulation skills with minimal human intervention
File(s)
Author(s)
Tsai, Ya-Yen
Type
Thesis
Abstract
Manipulation is vital in human daily activities. It makes interaction with an environment possible from simple tasks like grasping or writing to complex ones like assembly and surgery. Despite humans' versatility, there is continuous research in robotics to alleviate humans from monotonous or repetitive tasks. Presently, robotic manipulation is predominantly employed in well-structured and controlled industrial settings. As the interactions are not so complex, robots in these environments are typically programmed with task-specific code. However, applying robotic manipulation in everyday environments is considerably more intricate than in industrial settings. Relying solely on human programmers is impractical and hence, robots need to develop the ability to autonomously reason about the environment, learn new skills, and adapt to unfamiliar surroundings.
This work delves into a promising approach to robot learning that combines human priors with self-exploratory training. Human priors, such as demonstrations, offer valuable insights into tasks, effectively reducing the learning time for robots. Subsequently, self-exploratory training empowers robots to enhance their skills, enabling them to reason about environment states, adapt to new environments, surpass human demonstrations, and gain deeper insights into tasks through interactions with their surroundings. While providing significant human priors can expedite learning, it also runs the risk of introducing bias that restricts robots from surpassing human performance. Conversely, excessive self-exploratory training may prolong exploration time and impede the development of natural and human-like behaviours in robots. Therefore, this thesis aims to find a suitable balance between human involvement and robot learning. Specifically, we address critical questions concerning the selection and utilisation of human priors to facilitate effective learning. We also seek to determine the optimal types and quantities of human priors that should be provided, as well as how these priors should be integrated into the learning process to maximise the robot's capabilities.
This work delves into a promising approach to robot learning that combines human priors with self-exploratory training. Human priors, such as demonstrations, offer valuable insights into tasks, effectively reducing the learning time for robots. Subsequently, self-exploratory training empowers robots to enhance their skills, enabling them to reason about environment states, adapt to new environments, surpass human demonstrations, and gain deeper insights into tasks through interactions with their surroundings. While providing significant human priors can expedite learning, it also runs the risk of introducing bias that restricts robots from surpassing human performance. Conversely, excessive self-exploratory training may prolong exploration time and impede the development of natural and human-like behaviours in robots. Therefore, this thesis aims to find a suitable balance between human involvement and robot learning. Specifically, we address critical questions concerning the selection and utilisation of human priors to facilitate effective learning. We also seek to determine the optimal types and quantities of human priors that should be provided, as well as how these priors should be integrated into the learning process to maximise the robot's capabilities.
Version
Open Access
Date Issued
2022-11-18
Date Awarded
01/07/2024
License URL
Advisor
Johns, Edward
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)