Meituan Rider Path Prediction System (“Super Brain”) Deep Dive
PERSPECTIVE
Meituan’s rider path prediction system (commonly known as the “Super Brain”) is one of China’s largest real-time operations research optimization systems—dispatching millions of riders and tens of millions of orders daily, with millisecond-level decision latency requirements. 一、五层引擎 / Section I: The Five-Layer Engine 🔧 第一层：时空网格——把物理世界数字化 / Layer 1: Spatiotemporal Grid — Digitizing the Physical World
Meituan segments every city into “Honeycomb Grids” of approximately 500m × 500m. These are sliced chronologically into periods (morning, lunch, evening, night) to construct the Spatiotemporal Grid (ST-Grid). The update frequency is minute-level, dynamically capturing active rider counts, pending orders, and traffic congestion within each grid block.
Dimension
Granularity
Core Rationale & Motivation
Spatial
500m Grid
The delivery radius is generally within 3km; a 500m grid provides sufficient precision for localized dispatching.
Temporal
10-30 Min Slices
During the lunch rush, order density and delivery capacity fluctuate dramatically within a 10-minute window.
Merchant
One Node per Store
Tracks meal preparation speeds, signature dish cooking cycles, and historical real- time preparation baselines.
User
Precise Building Coords
Identifies specific community gates, exact building coordinates, and vertical impedance like elevator wait times. Meituan ‘Super Brain’ Technical & Sociological Analysis 1 / 6 🔧 第二层：时间预测——算准"还要等多久" / Layer 2: Time Prediction — Calculating “How Much Longer”
The core of path prediction is not “distance,” but time cost. Meituan utilizes an ETA (Estimated Time of Arrival) model based on LSTM + Attention to predict four critical timestamps. Inputs include: route length, traffic lights, left turns, real-time traffic, weather (rain/snow extends delivery times by 30%–60%), historical average speeds, and individual rider profiles (the system knows Rider Wang is 15% faster than Rider Li, yielding different ETA predictions for the same route).
Target Timestamp
Prediction Method & Input Features
Accuracy Window
Meal Prep Time
Merchant’s historical prep speed + active in-progress order volume + dish complexity
± 3-5 mins
Pickup Wait Time
Current queue length inside the store + number of riders already arrived
± 2 mins
Riding Time
Physical topology distance + real-time traffic + weather impedance + vertical elevator/stair friction
± 3-8 mins
Delivery Handover Time
Community access control protocols + target floor + door-to-door delivery requests
± 1-3 mins Meituan ‘Super Brain’ Technical & Sociological Analysis 2 / 6 🔧 第三层：订单合批——“AI拼单"比人拼得好 / Layer 3: Order Bundling — “AI Batching” Outperforms Humans
This is the most financially lucrative component of path prediction. A single delivery costs around ¥7–8, but bundling two orders pushes the total marginal cost to just ¥9–10. The algorithm must decide in milliseconds: To whom should this new order be assigned? Is it worth making a rider wait to complete a batch? Following its rollout, average daily deliveries per rider increased by 20-30%, while customer overtime rates dropped significantly. 实时决策树（简化版） / Real-Time Decision Tree (Simplified) A骑手在大望路，手里有1单去国贸 / Rider A is at Dawang Road, holding 1 order heading to Guomao.
direction (further east of Guomao).
and financial cost of an entire extra trip.
more minutes for prep + ride 2km extra.
remains in the user’s promised ETA?
Final Decision: If Overtime Risk < Safety Threshold → Execute Bundle; Else → Dispatch to Rider B.
“Riders on the Way Before You Order”
This represents the most predictive tier of the system. The underlying logic: The system anticipates an upcoming order and dispatches capacity nearby in advance.
- User Behavior Modeling: If you order from the same restaurant every day at 12:05, the algorithm identifies a 93% probability that you will place an order within the next 10 minutes.
- Pre-dispatching: At 11:55, a rider is quietly guided towards that specific merchant cluster.
- Instant Match: The moment you click order, the matching distance drops from 3km to under 300m, blending prep and transit perfectly. Data shows that predictive dispatching shortened peak- period delivery times by 3-4 minutes and decreased riders’ empty-riding rate by 12%. Meituan ‘Super Brain’ Technical & Sociological Analysis 3 / 6
“Optimization on the Fly”
This marks the fundamental difference between instant on-demand logistics and traditional static logistics. Every 5-10 seconds, the cycle loops to re-evaluate the global optimum for all active riders and unfulfilled orders.
Dynamic Real-World Scenario
System Real-Time Response & Adjustment
A rider is at a red light; an adjacent merchant releases a new order.
Calculates route overlap; if compatible and safely within ETA bounds → Appends order instantly
Sudden traffic accident or unexpected road closure causes severe congestion ahead.
Bypasses the bottleneck within milliseconds, pushing a fresh alternative route to the rider’s app.
User suddenly changes the address or becomes unreachable via phone.
Re-prioritizes the queue, commanding the system to temporarily skip this order and advance to the next task.
The rider takes a wrong turn due to poor visibility or unfamiliarity.
The app automatically triggers a route-correction popup, recalibrating ETAs for all subsequent milestones. Meituan ‘Super Brain’ Technical & Sociological Analysis 4 / 6
Perspective — This is Not Path Prediction, It is a Power Structure
On the surface, the system is computing “the shortest path.” However, through a deeper sociological lens, it defines: Who must deliver what, at what cost, to whom, and when— this is the ultimate manifestation of resource allocation power.
Total Precision Strips Riders of Bargaining Power The system monitors each rider’s exact physical limits (e.g., maximum daily order workload) while logging their compliance history (those who never reject a task). Far from AI developing bias, this is a natural behavioral reinforcement loop learned by the machine: compliant riders are given tighter, high-density order chains because the system knows they will execute them. The higher the rider’s predictability, the lower their leverage in labor negotiation.
User Experience is Merely a Marginal Variable in the Formula The deliberate “buffer time” embedded within the promised delivery windows constitutes the system’s margin of safety. The system is well aware that when it promises “delivery within 30 minutes,” a 25-minute delivery will completely satisfy the user. In crisis scenarios (e.g., storms or massive order spikes), the algorithm reclaims this buffer. The user’s experience drops from “delight” to “barely on time,” but as long as it avoids the complaint threshold, global throughput is optimized.
Merchants are Locked Inside an Algorithmic Enclosure Meal prep time prediction is fundamentally a digital credit rating issued by the AI to the merchant. Establishments with erratic or slow cooking speeds receive lower algorithmic scores, prompting the system to stealthily degrade their priority in capacity allocation (fewer assigned riders). To survive, merchants are forced to adjust their kitchen workflows. No explicit financial penalties are required; the algorithm itself acts as the most cold and effective disciplinary tool. Meituan ‘Super Brain’ Technical & Sociological Analysis 5 / 6 三、系统规模指标 / Section III: Scale & Operational Metrics
Core Operational Metric
Scale Data (Publicly Disclosed)
Daily Average Delivery Orders
~ 70 Million+ Orders
Real-Time Active Riders Online Daily
~ 1 Million+ Riders
Daily Path Re-routing Calculations
Billions of Computations
Core Decision Latency Cap
≤ 100 Milliseconds (ms)
Conclusion & Reflection Regarding the assertion that “human behavior can be precisely calculated,” Meituan’s path prediction delivers the ultimate proof: it is not that the algorithm accurately guesses which road you will take; rather, the algorithm makes that specific road your only viable choice. When the app instructs you to “please follow Route A,” you believe you are exercising choice, whereas the system has already completed the checkmate on the sovereign’s chessboard. Meituan ‘Super Brain’ Technical & Sociological Analysis 6 / 6