Reinforcement Learning for Mobile Robot Collision Avoidance in Navigation Tasks

<p>Collision avoidance is fundamental for mobile robot navigation. In general, its solutions include: {\it map-based} and {\it mapless approaches.} In the map-based approach, robots pre-plan collision-free paths based on an environment map and follow their paths during navigation. On the other hand, the mapless approach requires robots to avoid collisions without referencing to an environment map. This thesis first studies the map-based approach for multiple robots to collectively build environment maps. In this study, a robot following a pre-planned path may encounter unexpected obstacles, such as other moving robots and obstacles inaccurately presented on an environment map. This motivates us to study mapless collision avoidance in the second part of the thesis. Mapless collision avoidance requires a robot to infer an optimal action based on sensor data and operate in real time. Inferring an optimal action in a timely manner is computationally expensive, particularly when a robot has limited on-board computing resources. To avoid the expensive online action inferring, this thesis presents a reinforcement learning approach which learns policies for mapless collision avoidance under real-world settings. We first propose a Real-Time Actor-Critic Architecture (RTAC) to support asynchronous reinforcement learning under real-time constraint. Based on RTAC, we propose asynchronous reinforcement learning methods for mapless collision avoidance of various numbers of robots under different environment configurations. Through extensive experiments, we demonstrate that RTAC serves as a solid foundation to support multi-task and multi-agent learning for mapless collision avoidance under asynchronous settings.</p>

Name thesis:degree_name

Doctor of Philosophy (PhD)

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical Engineering and Computer Science

Year

2020

Author dc:creator

• Jiao, Zilong

Contributors dc:contributor

• Jae C. Oh

• Asynchronous Learning
• Collision Avoidance
• Multi-Agent Learning
• Multi-Task Learning
• Navigation
• Reinforcement Learning
• Engineering

Repository record dc:identifier

https://surface.syr.edu/etd/1227

OAI identifier oai:identifier

oai:surface.syr.edu:etd-2228