SVD-BDRL: A trustworthy autonomous driving decision framework based on sparse voxels and blockchain enhancement

With the rapid development of autonomous driving technology, issues regarding safe decision-making and multi-vehicle collaboration in complex urban environments have become increasingly prominent. To address the limitations of existing deep reinforcement learning methods in computational efficiency, decision transparency, and system safety, this paper proposes a novel framework, SVD-BDRL, which integrates a sparse voxel decoder and blockchain-enhanced deep reinforcement learning. The framework brings three key innovations: a sparse voxelization method to reduce computational complexity at the perception layer; a blockchain-based distributed experience management system to ensure data authenticity at the decision layer; and a real-time anomaly detection system combining graph neural networks and consortium blockchain for verification. Experimental results demonstrate that on the NuScenes and CARLA datasets, SVD-BDRL outperforms current methods, achieving an 11% reduction in collision rate and a 3.4% decrease in trajectory error, while maintaining real-time performance at 23.5 FPS. This study presents a promising new approach for creating safe, trustworthy autonomous driving systems, which is crucial for the commercialization of autonomous vehicles.