Autonomous Driving Infrastructure: Foundations, Key Trade‑offs, and Evolution Roadmap

Introduction – DiDi began autonomous‑driving research six years ago. The industry has experienced hype, capital cycles, and ongoing technical skepticism because the technology is still in its early stages. This article examines the basic architecture of autonomous driving, the main contradictions in its evolution, and the long‑term vision.

1. Basics of Autonomous‑Driving Infrastructure

The infrastructure consists of three layers: hardware, onboard (vehicle‑side) systems, and cloud services. Its core value is to provide a "nervous system" that tightly couples the software stack with the vehicle, ensuring safety and performance.

2. Main Technical Contradictions

• Rapid iteration vs. functional safety : Internet companies favor short iteration cycles, while automotive safety requires multi‑year cycles. The choice between Linux (fast iteration) and RTOS (safety) illustrates this tension.

• Sensors vs. compute power : Higher‑resolution cameras and denser LiDAR increase perception capability but demand more compute, power, and cooling, which are limited in a vehicle’s compact environment.

• Hardware performance vs. automotive‑grade safety : Server‑class x86 platforms offer high compute but lack automotive‑grade power, thermal, and reliability specifications. Even popular platforms like Nvidia Xavier (≈30 TOPS) fall short for full‑L4 autonomy.

3. Long‑Term Goal

The ultimate aim is safe, reliable driverless operation that removes humans from repetitive, dangerous driving tasks, improving traffic efficiency and reducing risk. Safety standards such as ISO 26262 define functional‑safety requirements, while newer approaches explore redundancy, V2X, and remote‑assist mechanisms.

4. Development Stages (based on MPI – miles per intervention)

Stage 1 (MPI < 10): Prototype exploration, heavy reliance on safety drivers, focus on rapid infra iteration.

Stage 2 (MPI < 100): Growth phase, balanced hardware upgrades and algorithmic research, more efficient simulation.

Stage 3 (MPI < 1000): Maturation, emphasis on system stability, safety, and hardware‑software co‑design.

Stage 4 (driverless): Full autonomous operation.

5. Evolution of the Infrastructure

Hardware – Sensors (LiDAR, high‑resolution cameras) move toward automotive‑grade specifications; compute platforms transition from generic x86 + Nvidia GPU to specialized, high‑efficiency, automotive‑qualified SoCs (e.g., Nvidia Pegasus, custom x86 solutions). Power, cooling, and interface considerations (12 V supply, liquid cooling, Ethernet/ CAN) are critical.

Software – Middleware sits between hardware and algorithms. Three representative stacks are ROS (general robotics), Apollo Cyber RT (high‑performance autonomous‑driving), and Iceoryx (high‑performance, safety‑oriented). The roadmap typically evolves from ROS/Ubuntu (flexibility) → Apollo‑style (performance) → safety‑certified middleware for production.

6. Conclusion

Autonomous driving is a complex, inter‑dependent system. Understanding the rhythm of development and aligning hardware, software, and safety efforts is essential for a sustainable roadmap.

Recruitment Notice – DiDi is hiring for positions in computer vision, planning, prediction, high‑definition maps, simulation, infrastructure, vehicle hardware, and testing. Applications can be sent to [email protected].