RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework

TL;DR

RAD-2 combines diffusion-based trajectory generation with RL-optimized reranking for autonomous driving motion planning. The framework achieves 56% collision rate reduction through temporal consistency in policy optimization and structured feedback mechanisms. Real-world deployment validates improved safety and smoothness in urban traffic scenarios.

• •Generator-discriminator framework pairs diffusion models for trajectory candidates with RL-based quality assessment
• •Novel policy optimization (TCGRPO) and on-policy generator updates reduce collision risk 56%
• •Bird's Eye View simulation enables efficient large-scale training with real-world validation

Generated with AI, which can make mistakes.