Meituan Autonomous Vehicle Engine: Architecture, Challenges, and Resource Optimization

The article introduces the concept of the "autonomous vehicle engine" and uses challenges faced in simulation environments as a thread to present the core design of Meituan's autonomous vehicle engine.

01 Introduction – Recent years have seen rapid development of autonomous driving technology. While road testing remains essential, the industry is increasingly turning to offline simulation to improve efficiency. Meituan’s team recognized that simulation and on‑vehicle testing differ greatly, especially in resource usage, and proposed the autonomous vehicle engine to isolate these differences so that functional modules can run unchanged in both scenarios.

02 Autonomous Vehicle Engine – The engine provides infrastructure at three layers: mechanism (communication, scheduling, data, configuration, monitoring), tools (debugging, visualization, performance tuning, evaluation), and module (a unified computation model and runtime). An architecture diagram (Fig. 1) shows core modules such as Perception, Localization, and Planning supported by the engine.

Challenges

Behavior Consistency – In simulation, results can vary with resource allocation or machine load, leading to “behavior inconsistency”. The engine addresses this by ensuring deterministic scheduling and a standard computation model for each module.

Scheduling – Various schedulers are defined: online scheduler (used on‑vehicle), replay scheduler (reproduces saved schedules), ideal scheduler (optimistic for simulation), and condition‑driven scheduler (balances resource usage). Figures 2‑4 illustrate ideal vs. actual timing and scheduler classifications.

Resource Utilization – The engine optimizes three main aspects:

Uneven data demand: Perception and Localization dominate data usage (>85%). Distributed deployment using a TensorFlow‑like graph model (Node/Module) enables scaling across machines.

Redundant computation: By separating Node (dependency) from Module (computation), results from heavy modules (e.g., Perception) can be cached and reused for multiple evaluation runs (Fig. 9).

GPU/CPU imbalance: The engine integrates model management, allowing predictions to run on remote model‑serving services, thus isolating GPU workloads from CPU clusters.

Conclusion – Meituan’s autonomous vehicle engine abstracts mechanisms and tools for functional modules, supporting both low‑latency on‑vehicle and high‑throughput simulation environments. Coupled with Meituan’s big‑data infrastructure, the engine aims to decouple module development from platform and environment, accelerating the deployment of autonomous delivery vehicles.