Terrain Cost Learning from Human Preferences for Robot Path Planning Using a Visual User Interface

<p>Robot navigation in terrains with limited exploration and limited knowledge has been a problem of interest in robotics due to the potential dangers that may arise during traversal. Due to the large number of path permutations within a complex and feature-rich real-world environment, and in the interest of saving time and ensuring safety, the robot should learn the optimal path without repeated exploration of the terrain. This can be accomplished by leveraging the path preferences of a human operator so that, with selective inputs, the agent can effectively learn a terrain-cost mapping in order to determine the optimal route, thereby eliminating the need for traversal.</p> <p>This thesis aims to achieve this goal by employing an inverse reinforcement learning (IRL) framework based on prior work in preference-based inverse reinforcement learning (PbIRL) [21] in conjunction with a visual user interface. In this work, the agent obtains relative preferences between two trajectories from a human and then decides which preference queries to make next using an active learning approach. The agent would then learn a final set of weights for terrain types within the environment, displaying the optimal path for traversal, after each set of queries.</p> <p>I validate my system by creating a situated visual user interface featuring a robot in a simulation environment featuring custom terrain types. By conducting experiments with real human subjects and a simulated user model, I demonstrate that the learning algorithm converges to a preferred path based on active learning of terrain costs from human preferences. With an average of 55.7 queries for a simulated user and 22.8 queries for human users, the algorithm converges to the preferred path across various terrain configurations.</p>

Name thesis:degree_name

M.S.

Level thesis:degree_level

Masters Thesis

Year dc:date.available

2023

Author dc:creator

• Velagapudi, Kaivalya

Contributors dc:contributor

• Christopher Reardon

• Robotics
• Navigation
• Path planning
• Artificial Intelligence and Robotics
• Computer Sciences
• Physical Sciences and Mathematics

Statement dc:rights

• <p>Copyright is held by the author. User is responsible for all copyright compliance.</p>

Language dc:language

en

Repository record dc:identifier

https://digitalcommons.du.edu/etd/2214

OAI identifier oai:identifier

oai:digitalcommons.du.edu:etd-3205