Talks and Poster Presentations (with Proceedings-Entry):
M. Kobelrausch, A. Jantsch:
"Collision-Free Deep Reinforcement Learning for Mobile Robots using Crash-Prevention Policy";
Talk: 2021 7th International Conference on Control, Automation and Robotics (ICCAR),
Singapore;
2021-04-23
- 2021-04-26; in: "2021 7th International Conference on Control, Automation and Robotics (ICCAR)",
IEEE,
(2021),
ISBN: 978-1-6654-4986-1;
52
- 59.
English abstract:
In this paper, we propose a crash-prevention policy for an autonomous collision-free mobile robot based on deep reinforcement learning. The objective is to reach a random location in a limited workspace safely. We go beyond the well-treated navigation paradigm by introducing a crash-prevention policy derived from action-sensor-space characteristics to achieve collision-free learning. This approach enables efficient and safe exploration by guaranteeing continuous collision-free actions, especially for agents learning in physical systems. We use Deep Deterministic Policy Gradient as a base method to evaluate the proposed crash-prevention policy on a mobile robot environment. Experiments show that using our approach maintains or even slightly improves training results while collisions are entirely avoided.
Keywords:
Safe Reinforcement Learning, Deep Reinforcement Learning, Continuous Control, Collision-Free Learning, Mobile Robots
"Official" electronic version of the publication (accessed through its Digital Object Identifier - DOI)
http://dx.doi.org/10.1109/ICCAR52225.2021.9463474
Created from the Publication Database of the Vienna University of Technology.