Advances in Machine Learning for Real‑Time Delivery at Meituan

Meituan’s AI team presents the latest progress of machine‑learning techniques applied to instant food delivery, aiming to perceive real‑world scenarios and predict every detail of the delivery process.

The rapid growth of instant delivery worldwide (e.g., Uber Eats, Deliveroo, Swiggy) creates new technical challenges that require senior experts in machine learning, operations research, LBS, and NLP.

Inspired by the sci‑fi concept of a “super brain”, Meituan built the “Superbrain” delivery system, which integrates big‑data processing, real‑time feature computation, and a suite of AI modules for perception, prediction, and decision‑making.

The core business model of instant delivery focuses on efficiency, cost, and user experience. It forms a positive feedback loop: higher efficiency reduces cost, which improves user experience, attracting more users and merchants, further boosting efficiency.

Key components of the Superbrain system include:

Big‑data processing and computing platforms (real‑time feature pipelines, offline data, ML platform).

World perception via LBS and multi‑sensor data (indoor positioning, fine‑grained scene description, rider motion detection).

Accurate understanding and prediction (ETA, demand, capacity forecasts).

Complex decision making (real‑time multi‑rider dispatch, dynamic pricing, network planning).

Machine‑learning challenges are high accuracy and granularity, handling noisy offline data (GPS drift, incomplete merchant data), and ensuring robustness to weather and traffic variations.

ETA (Estimated Time of Arrival) is a critical parameter linking user experience, cost, and downstream decisions. Meituan predicts ETA at building‑ and floor‑level by:

Precise address parsing to building/unit/floor (level 5+).

Cleaning and aggregating rider trajectory data.

Using tree‑based models and smoothing techniques to handle sparse data.

Providing ETA to dispatch and pricing modules.

Map accuracy is essential. Meituan builds a dedicated delivery‑map solution that ensures correct real‑time rider locations, precise merchant/user addresses, and reliable navigation, surpassing the needs of traditional logistics or ride‑hailing maps.

Location correction is performed by mining “delivery points” from rider sign‑in data. The pipeline includes address grouping, noise reduction, aggregation, and confidence scoring, tackling GPS drift, user selection errors, and POI sparsity.

Results show a significant reduction in rider delivery distance (>100 m cases drop sharply) and high‑precision mapping to unit‑door level.

Beyond outdoor positioning, Meituan pursues “scenario perception” to capture indoor activities using Wi‑Fi/Bluetooth geofencing and motion‑state recognition, enabling finer control of dispatch, pricing, and accountability.

In summary, Meituan’s AI‑driven delivery system processes tens of millions of orders daily, performs billions of path calculations per hour, and continuously improves ETA, dispatch, map optimization, dynamic pricing, and scenario awareness to shorten delivery times from one hour to around 30 minutes.