Yandex Go et Lavka : réseaux de neurones graphiques et ML orchestrent des millions de trajets et dark stores
Global RL-assisted dispatch and chained-trip matching reduced driver deadhead mileage and average passenger wait times by 15–22%, increasing driver hourly utilization without raising rider fares.
| Company & Sector | Яндекс Go и Яндекс Лавка — Mobilité et Q-Commerce |
|---|---|
| Models & AI Stack | CatBoost · Graph Neural Networks (GNN) · Reinforcement Learning · Bipartite Matching · YTsaurus · Real-Time Geo Hexagons |
| Chronologie | 2020–2022 — ML auto-replenishment in Yandex Lavka and GNN ETA prediction; 2023–2024 — RL hexagon value estimation and dynamic courier batching; 2025–2026 — LLM-powered driver and courier support assistants. |
| Verification | Vérifié 2026-10 · 3 primary sources |
Contexte et échelle de l'entreprise
The Yandex Go super-app integrates ride-hailing, 15-minute grocery delivery (Yandex Lavka), restaurant delivery (Yandex Eats), and last-mile logistics across hundreds of cities. Every day, the platform orchestrates millions of rides and deliveries under shifting traffic and weather conditions.
In quick commerce across 500+ dark stores, unit profitability hinges on two variables: SKU-level demand forecasting across 4,000+ items per neighborhood and courier order batching density.
In ride-hailing, the primary operational lever is minimizing deadhead mileage (driver pickup time and empty kilometers) while balancing spatio-temporal supply and demand across hexagonal city grids.
Problème et goulets d'étranglement
In ultra-fast grocery retail, forecasting errors on perishables (dairy, fresh bakery, ready-to-eat meals, produce) cause either costly food waste write-offs or peak-hour stockouts. Manual replenishment across hundreds of hyperlocal dark stores is impossible.
In ride-hailing and courier dispatch, greedy nearest-driver matching is globally suboptimal over even a 3–5 minute horizon: assigning a driver to a marginal pickup can starve an adjacent hexagon seconds before a demand surge.
Meanwhile, matching decisions must execute within milliseconds while road-graph edge weights shift every minute due to traffic signals, accidents, and weather.
Architecture et déploiement IA
Yandex Go runs a multi-stage ML dispatch pipeline. ETA prediction combines Graph Neural Networks (GNNs) over the road network with CatBoost models ingesting live GPS telemetry from thousands of active vehicles.
Order matching is solved as a bipartite global optimization problem augmented with Reinforcement Learning (RL) state-value functions: the matcher evaluates not only immediate pickup distance, but also the expected future value of the destination hexagon and seamless chained-ride dispatching before the current trip ends.
In Yandex Lavka, an automated replenishment engine powered by CatBoost and deep time-series models predicts hourly demand for every SKU × dark-store pair using weather, holidays, promotions, and shelf-life constraints—autonomously issuing supplier purchase orders.
“"City-scale dispatch optimization is like playing chess several moves ahead: the goal is not merely assigning the closest car now, but preserving fleet balance across every neighborhood 20 minutes into the future." — Yandex Go Engineering
Impact mesurable et économie
Global RL-assisted dispatch and chained-trip matching reduced driver deadhead mileage and average passenger wait times by 15–22%, increasing driver hourly utilization without raising rider fares.
In Yandex Lavka, ML replenishment keeps perishable waste below 1.5–2% of turnover—2x to 3x lower than traditional grocery retail—while sustaining 15-minute delivery SLAs.
Algorithmic order batching and dark-store pick-path optimization lifted courier deliveries per hour by over 25%.
Leçons pour l'industrie
Greedy nearest-neighbor dispatch loses to global bipartite matching with RL state-value lookahead in high-density cities.
In 15-minute grocery delivery, SKU × dark-store × hour forecast accuracy directly dictates unit profitability.
CatBoost paired with Graph Neural Network road embeddings remains the gold standard for low-latency spatio-temporal tabular prediction.
Chained trip assignment (matching the next ride 2 minutes before drop-off) unlocks fleet capacity without adding vehicles.
Sources primaires vérifiées
Questions clés sur cet article
Why is perishable food waste in Lavka several times lower than in traditional supermarkets?
Dark stores have no walk-in display damage, and the ML engine orders exact daily quantities for a 1.5 km radius factoring in weather and day-of-week, triggering automated markdowns before expiration.
How can a driver be matched while still completing a prior trip?
The GNN ETA engine predicts the exact drop-off second and coordinates, entering the vehicle into the matching pool for pickups near the destination.
Why are Graph Neural Networks needed for ETA prediction?
Congestion on one intersection cascades nonlinearly into adjacent streets. GNN message passing across road-graph edges captures traffic waves before static historical averages react.
Executive AI Consulting, Custom Agents & Team Workshops
Work directly with Zhemal Khamidun at hamidun.com — enterprise AI architecture, production RAG & multi-agent systems, and executive training.
L'essentiel de l'IA — une fois par semaine
Sept actus qui ont vraiment compté, choisies à la main. Sans bruit ni communiqués.
C'est fait ! Vérifiez votre boîte mail pour la confirmation.