D$^2$Drive: Training-Free Dynamic Inference for End-to-End Autonomous Driving

End-to-end autonomous driving frameworks have achieved strong performance in perception, prediction, and planning. However, their deep interaction architectures often introduce substantial computational overhead and inference latency, which hinders real-time deployment. To improve inference efficiency without retraining, we propose D$^2$Drive, a training-free dynamic inference framework for end-to-end autonomous driving.
D$^2$Drive reduces redundant computation by exploiting feature stability during inference. It contains two complementary components. First, a dynamic exit mechanism terminates deeper computation when feature differences between adjacent layers become sufficiently small. Second, an FFN partial computation strategy selectively performs FFN operations on tokens with larger variations across layers, while stable tokens bypass redundant FFN computation. These two designs reduce computation at both the layer and token-computation levels.
Unlike pruning, quantization, and retraining-based acceleration methods, D$^2$Drive does not modify model parameters, task heads, or training objectives. Therefore, it can be directly applied to pretrained autonomous driving frameworks during inference. We evaluate D$^2$Drive on four representative end-to-end autonomous driving frameworks, including UniAD, VAD, SparseDrive, and MomAD. The evaluation covers GFLOPs for the whole model, computation inside selected modules, planning accuracy, collision rate, tracking quality, and detection performance, providing a systematic protocol for analyzing the balance between efficiency and accuracy in training-free dynamic inference for autonomous driving systems.