Risk-constrained Spatio-temporal Attention Ppo for Lane-changing Decisions in Mixed Traffic
Year: 2026
Type: article
Abstract: Lane-changing decisions in mixed traffic must reconcile task completion, surrounding-vehicle interaction, and safety under partial observability. A vanilla Proximal Policy Optimization (PPO) policy that flattens neighboring-vehicle features tends to obscure which vehicles are relevant to the current maneuver, while directly adding a risk penalty to the reward does not provide an explicit budget for cumulative risk. This paper develops a reproducible three-lane Eclipse SUMO environment and proposes a risk-constrained spatio-temporal attention PPO method. The method represents the ego vehicle and up to eight neighboring vehicles as traffic-agent tokens, applies multi-head attention within each frame, aggregates four historical frames with a gated recurrent unit (GRU), and introduces a separate cost critic with a Lagrangian risk budget. The experimental protocol compares vanilla PPO, attention-enabled PPO, risk-constrained PPO, and the full model under identical traffic, reward, and optimizer settings across three training seeds and 20 evaluation episodes per seed. Across the resulting 240 evaluation episodes no policy produced a recorded collision, so the policies were separated by task completion and surrogate risk rather than by collision counts. The full model reached the target lane in 76.7% of its episodes with a mean episode risk cost of 0.040, compared with 46.7% and 0.247 for vanilla PPO, at a mean speed 0.42 m/s lower; attention and the risk constraint each improved completion modestly on their own, and only their combination produced a gain whose direction was consistent across all three seeds. The results support the modular design of the method, and they also show that the simulator safety backstop, rather than the collision count, sets the limit on what this evaluation can claim about safety.
DOI: 10.61784/wjit3127
Author: ZeYu Chen
Institution: China Fire and Rescue Institute
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214762295