Autonomous-driving planners are not judged by whether they draw plausible trajectories. They are judged by what happens after the car commits to one. RAD-2 is notable because it pushes diffusion-based planning deeper into closed-loop feedback instead of relying only on imitation learning.

The arXiv paper, submitted on 16 Apr 2026 at 17:59:44 UTC, starts from a practical weakness in high-level driving systems. Diffusion planners can model multiple possible futures, but the paper says they can suffer from stochastic instability and lack corrective negative feedback when trained purely from demonstrations.

RAD-2 answers that with a generator-discriminator framework. A diffusion generator produces diverse trajectory candidates. An RL-optimized discriminator then reranks those candidates according to long-term driving quality. That separation is the key engineering choice: instead of forcing sparse scalar rewards directly onto a high-dimensional trajectory space, RAD-2 uses the discriminator to turn closed-loop outcomes into a more manageable selection signal.

The paper adds two training pieces around that structure. Temporally Consistent Group Relative Policy Optimization is designed to make reinforcement learning use temporal coherence more effectively, while On-policy Generator Optimization converts closed-loop feedback into structured longitudinal optimization signals. The authors also introduce BEV-Warp, a high-throughput simulation environment that evaluates closed-loop behavior directly in Bird's-Eye View feature space through spatial warping.

The reported result is the reason this belongs on the watch list: RAD-2 reduces collision rate by 56% compared with strong diffusion-based planners. The authors also say real-world deployment improved perceived safety and driving smoothness in complex urban traffic, a stronger claim than an offline benchmark alone.

For the autonomous-driving stack, the interesting question is whether generator-discriminator planning becomes a practical way to keep the diversity benefits of diffusion while adding stronger correction from deployment-like feedback. The next thing to watch is how well BEV-Warp and the project artifacts hold up when other teams try to reproduce the closed-loop gains.

#planning

RAD-2 cuts autonomous-driving collisions by 56% in closed-loop tests