DIVER: Reinforced Diffusion Breaks Imitation Bottlenecks in End-to-End Autonomous Driving

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Existing end-to-end autonomous driving (E2E-AD) methods often exhibit conservative behaviors due to reliance on single-expert imitation learning.
• This homogeneity limits their ability to generalize to complex, real-world driving scenarios.

Purpose of the Study:

• To develop a novel E2E-AD framework, DIVER, capable of generating diverse, safe, and goal-directed driving trajectories.
• To address the mode collapse and generalization issues prevalent in single-mode imitation learning approaches.

Main Methods:

• DIVER utilizes diffusion-based multi-mode trajectory generation conditioned on map elements and surrounding agents.
• It employs Group Relative Policy Optimization (GPRO) to treat the diffusion process as a stochastic policy, optimizing for trajectory diversity and safety.
• A novel trajectory diversity metric is introduced to evaluate multi-mode predictions, overcoming limitations of L2-based metrics.

Main Results:

• DIVER significantly enhances trajectory diversity compared to traditional imitation learning methods.
• Experiments on NAVSIM, Bench2Drive, and nuScenes datasets demonstrate effective mitigation of the mode collapse problem.
• The framework produces physically plausible plans that encourage exploration beyond expert demonstrations.

Conclusions:

• DIVER offers a significant advancement in end-to-end autonomous driving by enabling diverse and safe trajectory generation.
• The proposed approach effectively tackles the limitations of single-mode imitation learning, improving generalization in complex scenarios.
• The combination of diffusion models and reinforcement learning, guided by GPRO, provides a robust framework for future autonomous driving research.