TY - JOUR
T1 - Comfort-Oriented Motion Planning for Automated Vehicles Using Deep Reinforcement Learning
AU - Rajesh, Nishant
AU - Zheng, Yanggu
AU - Shyrokau, Barys
PY - 2023
Y1 - 2023
N2 - Automated vehicles promise numerous advantages to their users. The proposed benefits could however be overshadowed by a rise in the susceptibility of passengers to motion sickness due to their engagement in non-driving tasks. Increasing attention is paid to designing vehicle motion to mitigate motion sickness. In this work, the deep reinforcement learning (DRL) method is used to plan vehicle trajectories, with a focus on minimizing low-frequency accelerations. These are known to be the primary cause of motion sickness. The goal is achieved by incorporating a frequency-weighted discomfort term into the reward function during training. The ability of the trained agent to target undesirable frequencies in accelerations is verified by comparing it with another agent trained for improving overall acceleration comfort. A reduction of 9.6% in frequency-weighted discomfort is achieved. The motion plan from the DRL agent is further compared with trajectories generated by human drivers in real-world scenarios. The results demonstrate comparable performance between the DRL agent and human drivers. Meanwhile, a significant reduction in online computation time has been observed when compared to a motion planner based on numerical optimization.
AB - Automated vehicles promise numerous advantages to their users. The proposed benefits could however be overshadowed by a rise in the susceptibility of passengers to motion sickness due to their engagement in non-driving tasks. Increasing attention is paid to designing vehicle motion to mitigate motion sickness. In this work, the deep reinforcement learning (DRL) method is used to plan vehicle trajectories, with a focus on minimizing low-frequency accelerations. These are known to be the primary cause of motion sickness. The goal is achieved by incorporating a frequency-weighted discomfort term into the reward function during training. The ability of the trained agent to target undesirable frequencies in accelerations is verified by comparing it with another agent trained for improving overall acceleration comfort. A reduction of 9.6% in frequency-weighted discomfort is achieved. The motion plan from the DRL agent is further compared with trajectories generated by human drivers in real-world scenarios. The results demonstrate comparable performance between the DRL agent and human drivers. Meanwhile, a significant reduction in online computation time has been observed when compared to a motion planner based on numerical optimization.
KW - Automated driving
KW - deep reinforcement learning
KW - motion planning
KW - motion sickness
KW - proximal policy optimization
UR - http://www.scopus.com/inward/record.url?scp=85161002854&partnerID=8YFLogxK
U2 - 10.1109/OJITS.2023.3275275
DO - 10.1109/OJITS.2023.3275275
M3 - Article
AN - SCOPUS:85161002854
SN - 2687-7813
VL - 4
SP - 348
EP - 359
JO - IEEE Open Journal of Intelligent Transportation Systems
JF - IEEE Open Journal of Intelligent Transportation Systems
ER -