Keywords: autonomous vehicles, vehicle safety, reinforcement learning, deep learning, decision making
Reinforcement learning (RL) is gaining more and more attention in autonomous driving as a promising decision making and control algorithm. However, it lacks fundamental assurance of safety in complex environment, in which the road vehicles operate. Namely, the agent must collide with other traffic participants to learn that collision must be avoided by any means necessary. Advanced driver assistance systems (ADAS), in turn, assist the driver in safety critical situations, hence helping to mitigate road accidents.
In this thesis a method of training the agent with RL, where traffic participants’ safety is addressed with ADAS functions, will be investigated. The agent must take into consideration foremost safety of the maneuver, as well as comfort and efficiency simultaneously. For instance, activation of the ADAS will guarantee vehicle’s safety signaling to the RL agent that a prior behavior must be avoided.
- Relevant literature analysis
- Design of the simulation environment in high fidelity driving simulator
- Design of the simplified ADAS and reward function for RL, which considers safety, comfort, and efficiency
- Train the RL agent in numerical simulation applying the designed method
Prerequisites: C++/python (high), reinforcement learning (medium), vehicle safety (low), Windows / Linux (medium)
Tools: pyTorch/tensorflow, Carla / LGSVL simulator, OpenAI Baselines or similar
Start: Available immediately
D. Chen, L. Jiang, Y. Wang, Z. Li, “Autonomous Driving using Safe Reinforcement Learning by Incorporating a Regret-based Human Lane-Changing Decision Model” 2020 American Control Conference (ACC), Denver, CO, USA, 2020.
X. Xiong, J. Wang, F. Zhang, K. Li, “ Combining Deep Reinforcement Learning and Safety Based Control for Autonomous Driving“, 2016, preprint arXiv: https://arxiv.org/abs/1612.00147