Dissertation Defense Announcement for Pradeep Chintam – 10/15/2024 at 2:00 PM

September 24, 2024

Dissertation Title: Sampling Space and Deep Reinforcement Learning Based Robot Navigation
When: 10/15/2024 2:00 PM
Where: Microsoft Teams Join the meeting now  Meeting ID: 233 724 710 995. Passcode: B9Jc5Y

Candidate: Pradeep Chintam
Degree: Doctor of Philosophy in Electrical and Computer Engineering
Committee Members: Dr. Chaomin Luo, Dr. John E. Ball, Dr. Yu Luo, Dr. Jenny Q. Du

Abstract:
Robot navigation through unpredictable complex environments presents a significant challenge within the realm of robotics, with path planning standing out as a pivotal element. Among the prevalent strategies for path planning, graph-based and sampling-based approaches emerge prominently. Rapidly-exploring Random Tree (RRT) is a sampling-based approach for solving the robot path planning problem. Although RRT discovers a path, it may not be optimal. RRT* discovers the optimal path asymptotically, but it has performance concerns. Informed RRT* and Batch Informed Trees (BIT*) enhance RRT* by synthesizing additional heuristic, Informed Sampling Space (ISS). In most real-world complex environments the heuristic, ISS, is trivial. Moreover, the uniform cost function used in these algorithms is not suitable for Informative Path Planning (IPP). In this research a new path biased sampling approach, Advised-RRT*, is developed. Advised-RRT* discovers the near-optimal path fast while preserving the probabilistic completeness and asymptotic optimality features of RRT*. Advised-RRT* finds the initial path utilizing Bi-directional RRT*, and then optimizes the path by biasing the sampling using a Targeted Advised Sampling Space (TASS), which is formed by overlapping n-spheres around the most recent solution path.

A new Informed Sampling Space (ISS) driven Informative Path Planning (IPP) approach is developed to facilitate autonomous robots to navigate and explore unknown and hazardous environments for in-situ resource utilization efficiently. The ISS-driven IPP approach is targeted on multi-objective optimization enabling the robot to plan its path and simultaneously explore multiple high-interest areas efficiently. A new cost function based on Multivariate normal (MVN) probability density function (PDF) and a normalization function is also developed in this research to incorporate the high-interest spots. Additionally, for the execution phase of a robot navigation, this research proposes a local motion planning approach using 2 steps. First, Advised RRT* is utilized for global path planning. Second, a Proximal Policy Optimization (PPO) based deep reinforcement learning (DRL) model is leveraged to navigate the robot through intermediate states. Simulation and comparative analysis substantiate the efficacy and robustness of the developed methodologies. Advised-RRT* has been evaluated using Moving AI benchmark and Open Motion Planning Library (OMPL). The OMPL comparison results, benchmark simulation results, T-Test results, and real robot experiment results validate that our Advised-RRT* algorithm is admirable to some other algorithms in terms of convergence rate and computational complexity. Also, the simulation results corroborate that our proposed ISS-driven IPP with RRT* converges rapidly towards the near-optimal solution with respect to both navigation time and environment exploration. Promising results were observed with DRL based local navigation during simulations.