New RL Framework Boosts Safe Autonomous Lane Changing.

Transfer learning is a promising approach for improving the efficiency of policy learning in autonomous systems by reusing knowledge from existing tasks. However, applying this to safety-critical domains like autonomous highway lane changing faces challenges due to distribution shifts between source and target environments, which can lead to unstable training and performance degradation. Furthermore, adapting to new target domains often requires extensive, potentially unsafe, exploratory interactions. To address these limitations, researchers have developed a new safe transfer reinforcement learning framework. This framework incorporates an adaptive teacher intervention mechanism that uses instantaneous safety costs to limit risky exploration, gradually reducing intervention as the system learns. This mechanism also generates dual-source samples for joint training. Additionally, the framework includes a teacher-guided safe transfer module that embeds action evaluation information from a teacher policy into the student's learning process through reward shaping, enhancing both safety and efficiency. A teacher-guided weighted optimization mechanism further stabilizes transfer performance by adjusting sample weights based on a likelihood ratio factor. Experimental results, including validation on the real-world NGSIM dataset, demonstrate that this method significantly outperforms baseline approaches in both safety and learning efficiency across various traffic conditions.