- ホーム
- > 洋書
- > ドイツ書
- > Mathematics, Sciences & Technology
- > Computer & Internet
- > hardware
Full Description
Build an end-to-end autonomous driving system that you can train, test, and put in the driver's seat of a simulator. This book takes you from research code for perception, prediction, mapping, and planning to a single learning pipeline you control, and shows you why each design decision is made, so you can adapt the system to your own vehicle, data, and safety requirements.
You'll start by turning published models into interchangeable modules and wiring them into one graph that runs a forward pass. From there, you'll follow the order real projects reward: label-free losses (rule-based, self-supervised, and distillation) let you shape the training pipeline before any dataset arrives, and nuScenes with supervised losses then fills it with labeled data. You'll train the model under distinct imitation learning regimes, see why a low validation loss does not mean a model can drive, and close the loop in the CARLA simulator with timestamped sensors and a compact driving automation system with safety pipes. Finally, you'll fine-tune the model with reinforcement learning.
Most material on end-to-end driving stops at a model architecture. This book is written from the perspective of an engineer who builds these systems for automotive clients, and it covers what papers leave out: contracts between modules, sensor fusion, configuration, and the safety logic around a learned model. Each chapter builds only on the code of the chapters before it, so you finish with a system you understand end to end, and a design you can keep extending as research moves on.
What You'll Learn
● Integrate models from research papers into a modular end-to-end driving pipeline with explicit contracts between tasks
● Configure, register, and test components with MMEngine, including custom CUDA extensions
● Design and test rule-based, self-supervised, and distillation losses before committing to a dataset
● Build nuScenes data pipelines with timestamped sensor bins and interchangeable data sources
● Train an end-to-end model under self-supervised, supervised, and distilled imitation learning regimes
● Connect the model to the CARLA simulator through a driving automation system with safety pipes
● Fine-tune the model with reinforcement learning
● Recognize where ASPICE and ISO 26262 affect how ML code is configured, versioned, and changed
Who Is This Book For
This book is for experienced machine learning and robotics engineers who want to build real end‑to‑end autonomous driving systems
