Loading...

Data-driven Methods for Cooperative Control of Wheeled Mobile Robots

Qahremani, Sina | 2021

475 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54370 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Sadati, Nasser
  7. Abstract:
  8. Employing wheeled mobile robots is growing in industry, transportation, space and defense industry and many other social fields as well. These robots are used to execute distinct forms of operations and tasks such as exploring the surface of the earth and other planets, serving in public places, backing natural disasters and warehousing, and so forth. In some cases, the assigned mission may not be capable of being performed as intended by a single robot. In this case, several robots will work together to execute a particular mission. Several research topics that are under investigation currently include the interacting procedure of robots as a multi-agent system in order to perform the assigned task, how the robots are located beside each other, and the process of path planning and not colliding with obstacles.Developing safe and efficient policies in decentralized scenarios in which each robot finds its own paths using limited and partial observations of other robots' states, and given information to avoid collisions in multi-robot systems will be considered as one of the challenges for the mentioned area. Collision avoidance with fixed and dynamic obstacles in multi-robot systems often requires constant communication between robots for local programming, which is not robust and is computationally limiting. Furthermore, the performance of these methods in practice is not comparable to centralized methods. In this thesis, a decentralized sensor-level collision-avoidance policy is developed for multi-robot systems, which has provided promising results compared to other methods. In the first step, for reducing the performance gap between decentralized and centralized methods, we have proposed a brand new method of training optimal policy in policy-gradient reinforcement learning field, which is multi-scenario and multi-stage. This policy employs a reinforcement learning algorithm to train a large number of robots in several complex environments. Ensemble Inference algorithm is used for improving the robustness and effectiveness of the trained policy during the test. In conclusion, the collision-avoidance policy in cooperation with deep reinforcement learning (DRL) can be considered as a significant solution for safe and effective routing of a multi-robot system in environments with fixed and dynamic obstacles
  9. Keywords:
  10. Deep Reinforcement Learning ; Collision Avoidance ; Reinforcement Learning ; Wheeled Mobile Robot ; Multirobot Systems ; Gradient Descent Algorithm ; Ensemble Inference Algorithm

 Digital Object List

 Bookmark