Local planning is a core component of the autonomous vehicles motion planning module. Various techniques were used to solve this problem, from classical approaches to machine learning ones. Challenges in this problem include solving the local planning problem efficiently while keeping high vehicle safety in various driving scenarios. In this paper, we present a local planning approach as a fusion between the A* search-based Supervised Learning (SL) model and the Deep Deterministic Policy Gradient (DDPG) based Reinforcement Learning (RL) model. The resulting Supervised-Reinforcement Learning (SRL) approach is compared to each of the separate approaches, in terms of training time and model accuracy. The results show that our proposed approach achieves better performance than pure RL in terms of training time and better than pure SL in terms of accuracy and safety while facing unseen scenarios.
Supervised-Reinforcement Learning (SRL) Approach for Efficient Modular Path Planning
08.10.2022
2214942 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Parking path planning method based on reinforcement learning
Europäisches Patentamt | 2020
|Intention-driven reinforcement learning-based path planning method
Europäisches Patentamt | 2024
|