Publication Date

Fall 2024

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Jun Liu; Mahima Suresh; Stas Tiomkin

Abstract

Numerous studies in robotics focus on enabling agents to perform tasks resembling human actions using their perception and other senses. Much of the research adopts the approach for deep reinforcement learning and feature engineering that performs effectively with controlled surroundings and well-thought reward signals for the agent to communicate with the environment. However, this creates dependence on the environment and requires a lot of domain expertise and supervision from the researcher. As a result, performance deteriorates when the agent’s surroundings change. To create a genuinely adaptive solution, agents should be open to and able to learn new skills based on the sparse rewards, with little or no supervision from the human, by directly interacting with the environment. This thesis presents an intrinsic motivation model for the agents to interact with and learn from their environment using visual mode of perception. Using intrinsic motivation, the agent performs well in the pre-defined environment, generates new goals, and learns new skills in different environments. Prior research in intrinsic motivation focused on proprioceptive information, and has successfully learned diverse skills with sparse rewards and minimal human supervision. However, humans interact with their surroundings using their eyes, like deciding the position of the cup for picking, or how to handle an egg. With visual modularity, the robotic agent can learn diverse skills more effectively by integrating information with vision. This thesis will discuss the impact of using visual sensor for perception, along with the proprioceptive information from an agent.

Download

Available for download on Wednesday, February 24, 2027

Share

COinS