NTR: Neural Token Reconstruction for Scene Token Bottleneck in End-to-End Driving
Researchers have developed Neural Token Reconstruction (NTR), a new framework designed to improve the scene token bottleneck in end-to-end autonomous driving systems. NTR uses a self-distillation masked latent reconstruction objective to ensure that compact scene tokens retain richer visual information for planning. This method, which removes auxiliary components at inference, has achieved state-of-the-art results on multiple autonomous driving benchmarks, including Waymo E2E and NavSim. AI
IMPACT Improves representation learning in autonomous driving models, potentially leading to more robust planning and decision-making.