[2304.10891] Transformer-Based mostly Autonomous Driving Fashions and Deployment-Oriented Compression: A Survey

View a PDF of the paper titled Transformer-Based mostly Autonomous Driving Fashions and Deployment-Oriented Compression: A Survey, by Juan Zhong and three different authors

View PDF
HTML (experimental)

Summary:Transformer-based fashions have gotten a central paradigm in autonomous driving as a result of they will seize long-range spatial dependencies, multi-agent interactions, and multimodal context throughout notion, prediction, and planning. On the identical time, their deployment in actual automobiles stays tough as a result of high-capacity attention-based architectures impose substantial latency, reminiscence, and vitality overhead. This survey evaluations consultant Transformer-based autonomous driving fashions and organizes them by activity function, sensing configuration, and architectural design. Extra importantly, it examines these fashions from a deployment-oriented perspective and analyzes how effectivity constraints reshape mannequin design decisions in apply. We additional assessment compression and acceleration methods related to Transformer-based driving programs, together with quantization, pruning, data distillation, low-rank approximation, and environment friendly consideration, and talk about their advantages, limitations, and task-dependent applicability. Reasonably than treating compression as an remoted post-processing step, we spotlight it as a system-level design consideration that immediately impacts deployability, robustness, and security. Lastly, we determine open challenges and future analysis instructions towards standardized, safety-aware, and hardware-conscious analysis of environment friendly autonomous driving programs.