Loading...

Optimal Path Planning of Autonomous Robots in Unknown Environments Based on Deep Reinforcement Learning

Estarki, Mohammad Reza | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57315 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Khodaygan, Saeed
  7. Abstract:
  8. In the present era, with the remarkable advancements in the robotics industry and the extensive development they have achieved in human societies, the role of robots is very fundamental and vital. Autonomous robots, as an essential part of this industry, are increasingly advancing and improving. One of the most crucial aspects of the efficiency of autonomous robots is their intelligent path planning and navigation. Although previous works have addressed the issue of path planning and navigation based on deep reinforcement learning, many challenges remain, especially in scenarios where the robot's knowledge of the environment is limited and minimal. In this research, by defining a deep reinforcement learning system, efforts have been made to improve existing algorithms and propose innovative structures, such as a non-identical critic network, for use in the improved twin delayed deep deterministic policy gradient algorithm, with the aim of enhancing the training process and making richer use of available information and data. This study defines a comprehensive and unique reward function to optimize a multi-objective problem and introduces an improved deep reinforcement learning algorithm. This research focuses on the path planning and navigation of autonomous robots in unknown and so-called mapless environments. The main objective is to achieve an optimal policy for single and multi-agent path planning in uncertain environments based on multi-objective optimization, such as the shortest safe path, smooth and fast movement towards goals. Therefore, various algorithms have been implemented in the prepared simulation infrastructure to empirically demonstrate the improvements made in the proposed algorithm. Additionally, in the case study section, the output algorithm is examined in a practical application of intelligent navigation, specifically in search and rescue missions. The idea is to provide a multi-section system in the form of search and navigation in an unfamiliar environment without having information about the locations of potential targets, so that if the robot detects individuals in danger during the search, it moves towards them. Focusing on the core of the research, various structures for actor-critic networks and different algorithms have been developed, which after training in increasingly complex single and multi-agent environments, have achieved collision rates of less than 5.5% for single-agent environments and 11% for multi-agent environments. Furthermore, by evaluating the trained model in test environments to measure its generalization ability in new and challenging scenarios, with a reduction of less than 6% in success rate, it is proven that an optimal policy has been achieved. The models' capabilities in crowded environments have also been assessed by increasing random obstacles in a simple environment, and the results indicate that the intelligent model is satisfactorily generalizable with increased environmental complexity
  9. Keywords:
  10. Autonomous Robot ; Deep Reinforcement Learning ; Actor-Critic Model ; Object Detection ; Search and Rescue Mission ; Non-Identical Critic Networks

 Digital Object List

 Bookmark

No TOC