Research on Autonomous Driving Decision-making Method Based on Inverse Reinforcement Imitation Learning

Le résumé fourni par la source

In response to the challenges of low sample efficiency and unstable training in automatic driving behavioral decision-making models, we propose a novel approach that combines behavioral cloning with generative adversarial imitation learning. By incorporating a time-weighting function into the loss function, we dynamically adjust the learning focus to facilitate smooth and orderly model training. Initially, greater emphasis is placed on behavioral cloning to establish a stable baseline decision strategy. Subsequently, through simulated annealing techniques, the weight assigned to behavioral cloning is gradually decreased while increasing the weight for generative adversarial imitation learning. This allows the model to acquire more nuanced decision features and potential reward functions derived from expert behavior. By engaging with real-world environments, the model further refines its decision-making capabilities and enhances its generalization ability in previously unencountered situations. The optimized decision strategies are integrated with fundamental rules to enable effective decision-making for autonomous vehicles. Finally, this approach was evaluated using the Carla car simulator. Experimental results indicate that under model control, the collision-free completion rate achieved 95.67%, demonstrating an effective method for developing a safe and reliable self-driving system.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.