This chapter consider another important application of machine learning to robotics, i.e., the utilisation of deep reinforcement learning agent for robot motion planning and control. We will first present some preliminaries on training DRL policy in robotics, followed by the discussion on the sample efficiency concerning the sufficiency of training (Sect. 13.3) and the introduction of several statistical methods for evaluation (Sect. 13.4). Afterwards, we will discuss how to formally express the properties (Sect. 13.5) and then focus on reusing the verification tools for convolutional neural network to work with deep reinforcement learning, by considering the verification of policy generalisation (Sect. 13.6), the verification of state-based policy robustness (Sect. 13.7), and the verification of temporal policy robustness (Sect. 13.8). In addition, we will discuss how to address the well-known Sim-to-Real challenge in robotics with the verification techniques (Sect. 13.9).
Deep Reinforcement Learning
Artificial Intelligence: Foundations, Theory, and Algorithms
2012-02-24
17 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
BASE | 2019
|Autonomous Driving with Deep Reinforcement Learning
SLUB | 2023
|