Most end-to-end autonomous driving systems stop improving once deployed. The new EvoAD framework, described in an arXiv preprint, aims to change that by adding a post-hoc reflection loop: after a trip, the system diagnoses how its trajectories could have been improved and carries those lessons into similar scenes later.
EvoAD stores driving experiences in a bird's-eye-view feature memory pool flagged by geometric and semantic similarity. Rather than using a world model only to simulate actions in advance, it applies a latent world model to evaluate candidate trajectories in a compact planning-oriented space. That lets the system turn executed outcomes into reusable improvement signals without retraining the base driving policy.
The authors report consistent gains across reflection rounds on the NAVSIM v1 and v2 benchmarks, with state-of-the-art results. They argue this points to world models serving as reflective mechanisms for post-deployment self-evolution in autonomous driving.