A Multi-Objective Real-Time Trajectory Planning Framework for Human–Machine Mixed Traffic Based on Self-Attention Guided CNN-LSTM

Human-machine mixed traffic (HMMT) where autonomous ground vehicles (AGVs) coexist with human-driven vehicles (HDVs), is expected to be the predominant transportation mode in the foreseeable future. To address the multi-objective real-time lane-change trajectory planning problem for AGVs in HMMT, this study proposes a self-attention guided convolutional neural network-long short-term memory network (CNN-LSTM) based framework. This framework introduces an interval type-2 fuzzy physical programming (IT2FPP) method to iteratively solve the trajectory planning problem under varying HDV driving intentions and constructs a dataset using the corresponding HMMT system states and AGV control actions. IT2FPP can handle uncertainties in all boundaries of the preference function and accommodate fuzzy preferences from multiple decision-makers, overcoming the limitations of type-1 fuzzy physical programming. Additionally, a fitting function-based obstacle avoidance method is proposed to model obstacles in the trajectory planning problem with fitting functions, ensuring collision-free trajectories while improving computational efficiency. Then, a self-attention guided CNN-LSTM network is designed to learn the mapping function between the HMMT system states and AGV control actions, enabling real-time trajectory planning with very low computational burden. The network effectively extracts intrinsic features from time series data, demonstrating high accuracy. As HDV driving intentions are incorporated in the dataset, there is no need for inference when generating AGV trajectories in real-time, further enhancing planning efficiency. The framework effectively addresses the multi-objective real-time lane-change trajectory planning problem for AGVs in HMMT, demonstrating high accuracy, real-time performance, and practical application potential, as validated by simulation and physical experiments.