DiDi Autonomous Driving Simulation System: Architecture, Cloud Platform, and Data Management

DiDi’s autonomous driving team, founded in 2016, has built a comprehensive simulation system to address the safety, cost, timeliness, and scale challenges of road testing. The system enables low‑cost, large‑scale replay of historical driving scenarios and the creation of fully virtual scenes to evaluate algorithm performance in extreme conditions.

The system is divided into three tightly integrated components:

Simulator: Simulates vehicle‑environment interactions in a single scenario and evaluates vehicle behavior with high fidelity. Techniques such as a virtual clock, timing simulation, and neural‑net caching are used to reproduce real‑world timing and improve performance, even on heterogeneous hardware.

Cloud Platform: Provides massive parallel compute resources to run thousands of scenarios concurrently. The platform leverages data‑parallelism of scenarios, virtual‑clock independence, and dynamic resource scheduling to achieve linear scalability with the number of worker nodes.

Data System: Manages the large volume of input and output data, offering scene‑library management, automated indexing and retrieval of road‑test data, and automated scoring, aggregation, and visualization of simulation results.

Key innovations include:

Running the simulator in both development containers and on actual vehicle‑server hardware to obtain accurate performance metrics.

Neural‑net caching that hashes model inputs to reuse inference results, yielding speedups beyond GPU‑accelerated inference.

Mixed‑deployment of cloud nodes on existing DiDi server resources, maximizing utilization during low‑traffic periods and minimizing cost.

The paper concludes with a set of requirements for future simulation systems: high‑fidelity functional testing, comprehensive scenario coverage for the Operational Design Domain (ODD), precise performance evaluation, and continued improvement of cloud resource efficiency. Meeting these goals is essential for accelerating the development and safe deployment of Level‑4+ autonomous driving technology.