Paper Overview
Field: Computer Vision Authors: Hao Gao, Shaoyu Chen, Yifan Zhu Published: 2025-04-17 arXiv: 2504.13094
Summary
High-level autonomous driving requires motion planners capable of modeling multimodal future uncertainties while remaining robust in closed-loop interactions. Although diffusion-based planners are effective at modeling complex trajectory distributions, they often suffer from stochastic instabilities and the lack of corrective negative feedback when trained purely with imitation learning.
To address these issues, the authors propose RAD-2, a unified generator-discriminator framework for closed-loop planning:
- A diffusion-based generator produces diverse trajectory candidates.
- An RL-optimized discriminator reranks these candidates according to their long-term driving quality.
- Temporally Consistent Group Relative Policy Optimization (TC-GRPO): leverages temporal coherence to mitigate the credit assignment problem in RL training.
- On-policy Generator Optimization: converts closed-loop feedback into structured longitudinal optimization signals, progressively guiding the generator toward high-reward trajectory manifolds.
- BEV-Warp: a high-throughput simulation environment that performs closed-loop evaluation directly in bird's-eye-view (BEV) feature space via spatial warping, enabling efficient large-scale training.
- RAD-2 reduces the collision rate by 56% compared to strong diffusion-based planners.
- Real-world deployment demonstrates improved perceptual safety and driving smoothness in complex urban traffic.
- Paper: https://arxiv.org/abs/2504.13094
This decoupled design avoids directly applying sparse scalar rewards to the full high-dimensional trajectory space, improving optimization stability.
Key Techniques
Results
Links
*Auto-collected on 2026-04-18*