Safe Reinforcement Learning using Data-Driven Predictive Control

被引:1
作者
Selim, Mahmoud [1 ]
Alanwar, Amr [2 ]
El-Kharashi, M. Watheq [1 ]
Abbas, Hazem M. [1 ]
Johansson, Karl H. [3 ]
机构
[1] Ain Shams Univ, Cairo, Egypt
[2] Jacobs Univ, Bremen, Germany
[3] KTH Royal Inst Technol, Stockholm, Sweden
来源
2022 5TH INTERNATIONAL CONFERENCE ON COMMUNICATIONS, SIGNAL PROCESSING, AND THEIR APPLICATIONS (ICCSPA) | 2022年
关键词
Reinforcement learning; robot safety; task and motion planning;
D O I
10.1109/ICCSPA55860.2022.10018994
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Reinforcement learning (RL) algorithms can achieve state-of-the-art performance in decision-making and continuous control tasks. However, applying RL algorithms on safety-critical systems still needs to be well justified due to the exploration nature of many RL algorithms, especially when the model of the robot and the environment are unknown. To address this challenge, we propose a data-driven safety layer that acts as a filter for unsafe actions. The safety layer uses a data-driven predictive controller to enforce safety guarantees for RL policies during training and after deployment. The RL agent proposes an action that is verified by computing the data-driven reachability analysis. If there is an intersection between the reachable set of the robot using the proposed action, we call the data-driven predictive controller to find the closest safe action to the proposed unsafe action. The safety layer penalizes the RL agent if the proposed action is unsafe and replaces it with the closest safe one. In the simulation, we show that our method outperforms state-of-the-art safe RL methods on the robotics navigation problem for a Turtlebot 3 in Gazebo and a quadrotor in Unreal Engine 4 (UE4).
引用
收藏
页数:6
相关论文
共 42 条
[1]  
Abbeel P., 2005, P 22 INT C MACH LEAR, P1, DOI DOI 10.1145/1102351.1102352
[2]  
Achiam J, 2017, PR MACH LEARN RES, V70
[3]  
Alanwar A., 2022, European Journal of Control, P100666
[4]  
Alanwar A, 2021, PR MACH LEARN RES, V144
[5]  
Alanwar A, 2022, Arxiv, DOI arXiv:2105.07229
[6]  
Althoff M., 2010, Reachability analysis and its application to the safety assessment of autonomous cars
[7]  
Ammar H. B., 2015, SAFE POLICY SEARCH L
[8]  
Asadi K., 2019, Combating the compounding-error problem with a multi-step model
[9]  
Hamrick JB, 2021, Arxiv, DOI [arXiv:2011.04021, 10.48550/arXiv.2011.04021]
[10]   Ensuring Safety, Security, and Sustainability of Mission-Critical Cyber-Physical Systems [J].
Banerjee, Ayan ;
Venkatasubramanian, Krishna K. ;
Mukherjee, Tridib ;
Gupta, Sandeep Kumar S. .
PROCEEDINGS OF THE IEEE, 2012, 100 (01) :283-299