Neural MPC-based Decision-making Framework for Autonomous Driving in Multi-Lane Roundabout

Yao Mu, Zhiqian Lan, Chang Liu, Ping Luo, Shengbo Eben Li

TL;DR: We propose a neural model predictive control algorithm which learns a dynamics model with interaction data and solves the optimization problem with the gradient guidance via model predictive control.

Abstract: Roundabouts are a popular alternative to traditional intersections due to their ability to reduce traffic collisions, but they pose challenges for autonomous vehicles due to their complex traffic situation. Previous attempts at decision-making in multi-lane roundabouts have been hindered due to the difficulty of accurately predicting the behavior of other traffic participants and the complexity of the overall traffic flow. To address the above challenges, this paper proposes a neural MPC-based decision-making framework for autonomous vehicles, in which multiple backup static paths are generate in real-time according to the road topology, and the decision-making problem is formulated as a series of parallel static path tracking problems with safety constraints, subject to dynamic surrounding vehicles. To overcome the uncertainty of traffic dynamics, we propose a neural model predictive control algorithm (NMPC), which learns a dynamics model with interaction data and solves the optimization problem with the gradient guidance via model predictive control. The path with the lowest cost is then chosen as the target path after solving all the constrained tracking problems and the corresponding action is chosen accordingly. To enhance computational efficiency, a critic network is used to approximate the constrained tracking cost and an actor network to approximate the control policy, reducing the burden of online solving. To evaluate the proposed framework, a multi-lane roundabout simulator is built to benchmark a real roundabout in Beijing and the proposed approach is tested with various densities of traffic flow. The results show that our method can successfully navigate the roundabout, perform lane change maneuvers safely and efficiently.

Track: Original Contribution

Type: PDF