IRL-DAL: Safe and Adaptive Trajectory Planning for Autonomous Driving via Energy-Guided Diffusion Models

IRL-DAL Framework Enhances Autonomous Vehicle Navigation

A new inverse reinforcement learning framework named IRL-DAL is set to advance the safety and adaptability of trajectory planning in autonomous vehicles. The system achieves a 96% success rate and dramatically reduces collision instances to just 0.05 per 1,000 steps.

The IRL-DAL framework begins training through imitation of an expert finite state machine (FSM) controller, integrating environmental feedback with an inverse reinforcement learning (IRL) discriminator signal. Following this phase, the system undergoes reinforcement learning (RL) utilizing a hybrid reward structure that combines environmental insights with targeted IRL incentives.

A crucial component is a conditional diffusion model, which acts as a safety supervisor for planning safe driving paths and maintaining lane discipline. Additionally, a learnable adaptive mask (LAM) enhances the vehicle's perception by adjusting visual attention based on speed and proximity to hazards.

After the initial imitation phase, the driving policy is refined using Proximal Policy Optimization (PPO), with training occurring within the Webots simulator.

IRL-DAL adeptly navigates unsafe environments, demonstrating expert-level handling of complex driving conditions. The research team has made the framework's code publicly available to encourage further development in autonomous driving.