AWGAIL: Auxiliary Human Decision-Making with Wasserstein GAIL for End-to-End Autonomous Driving

In autonomous driving decision-making, end-to-end approaches often rely on reinforcement learning (RL) or imitation learning (IL) to map raw sensory inputs directly to control actions. Generative adversarial imitation learning (GAIL) has recently emerged as a powerful IL method, but its reliance on Jensen-Shannon divergence can lead to training instability and poor robustness when expert demonstrations are limited or unbalanced. To address these issues, we propose auxiliary wasserstein generative adversarial imitation learning (AWGAIL). By incorporating a Wasserstein-based adversarial objective and auxiliary human decision patterns, AWGAIL enhances training stability and effectively mitigates sensitivity to abnormal demonstration data. Experimental results using the CARLA simulator show that AWGAIL surpasses existing methods in multiple driving scenarios, delivering more robust, reliable, and efficient autonomous driving performance.