| Reinforcement-Tracking: An End-to-End Trajectory Tracking Method Based on Self-Attention Mechanism
|
| Guanglei Zhao 1,2, Zihao Chen 1,2, Weiming Liao 1,2 |
1Institute of Electrical Engineering , Yanshan University
2Intelligent Rehabilitation and NeuroRegulation Key Laboratory of Hebei , Yanshan University |
|
|
|
|
|
|
|
| ABSTRACT |
|
As a bridge between planners and actuators, trajectory tracking is a critical and essential part of robot navigation, autonomous driving and some other fields. While traditional trajectory tracking methods generally have disadvantages such as poor tracking accuracy, high modeling requirements, and heavy computational load. This paper proposes an end-to-end trajectory tracking method based on reinforcement learning, and an information encoding network and a reinforcement learning policy network are constructed. A multi-task dense reward function for trajectory tracking is designed. For efficient encoding of local trajectory information, a self-attentive mechanism is developed. A virtual simulation environment is constructed for model training by modeling the trajectory tracking task. Comparing with model predictive control and pure pursuit in the tracking experiments of several reference trajectories, the obtained results show that the proposed method has significant advantages in terms of lateral tracking.
|
| Key Words:
Reinforcement learning · Trajectory tracking · Self-attention · Reward function
|
|
Corresponding Author.
Guanglei Zhao , Email. 
Preview 
Full text via DOI 
Download Citation
Print
