Quasi‐Continuous Decision‐Making for Autonomous Driving Based on a Hybrid Data‐Driven Approach

ABSTRACT Behaviour decision‐making technology is a crucial component of autonomous driving systems. Nowadays, artificial intelligence‐based behaviour decision‐making methods still have some problems, including unreasonable behaviours caused by discontinuous behaviour modelling and the respective limitations of supervised learning and reinforcement learning schemes. To address the above issues, a quasi‐continuous decision‐making method for autonomous driving is proposed. In terms of mathematical modelling, a large amount of manual driving data is collected, which makes a list of basic trajectories as the behaviour space for autonomous driving. The Euclidean distance from any element in the trajectory list to the features of the reference trajectory element is used to construct a list index and taken as a random variable to achieve a stochastic representation of the behaviour in the autonomous driving. In terms of the model‐solving scheme, a hybrid data‐driven approach is used to fit the optimal behaviour probability distribution under any state. The weighted JS‐divergence between the probability density of the manual driving strategy and of the environmental interaction strategy is used as the loss function unit to realise the hybrid driving of manual driving data and environmental interaction data. In simulation and field tests, the proposed scheme achieves an improvement of at least 12% in behavioural continuity compared to other baseline methods. The travelling efficiency reaches more than 80% of the benchmark level of the traffic environment, and occupant subjective evaluations are favourable.