Russian Post: OCR and robots sort around 8 million mail items a day

• PostTech: как устроена OCR-сортировка писем (5 проекций, 40 000 писем/час, SMAB, Сортмастер, видеокодирование, АСЦ Внуково) — Habr (блог PostTech / Почта России), 2020
• Что умеет Почта России и какие технологии развивает (объёмы, робот-манипулятор 4 сек, нейросети для маршрутов, свёрнутые роботы-курьеры) — Hi-Tech Mail.ru, n/a
• «Почта России» внедряет систему моделирования маршрутов с ИИ («Сколково» + «Тераплан», 2024–2026; 54 млн км маршрутов в месяц) — CNews, 2024-05-28
• Почта России внедрила ИИ-систему для оптимизации логистики (платформа «Тераплан», до 30 млн операций ежедневно) — Online47, 2026-04-08
• Почта России: проекты искусственного интеллекта (голосовой ассистент на Yandex SpeechKit, 92–95%) — TAdviser, n/a
• Динамика доставки в срок 2013–2016: письма 54% → 85%, посылки 52% → 95% — РБК Тренды, 2022-01
• Генеральный директор АО «Почта России» — Михаил Волков — Почта России, n/a
• Главой Почты России стал Михаил Волков — Forbes Russia, 2023-01

Background

Russian Post is one of the country's largest logistics systems: about 38,000 post offices, 311,000 employees, around 2.6 billion letters and about 400 million parcels a year — roughly 8 million mail items per day (Hi-Tech Mail.ru). The flip side of scale is harsh economics: paper mail volumes decline year over year (2024 materials already cite 1.3 billion letters and 240 million parcels annually — CNews), and every manual sorting operation weighs on unit costs.

The company's sorting network comprises 1,100 nodes of different tiers: macro-regional, regional, and local. The flagship is the automated sorting center in Vnukovo — 64,000 m², capable of handling 3 million items daily; the equipment came from Toshiba, Siemens, and Vanderlande, while the software layer around it was written in-house by the PostTech team. When the technical write-up was published (2020), the company planned 36 more logistics centers by 2022 — sorting was deliberately concentrated in large 'factories' where automation pays off through volume. Mikhail Volkov has been the company's CEO since January 2023.

Sorting automation is part of a years-long logistics modernization that began well before the AI hype: back in 2013–2016 the share of letters delivered on time grew from 54% to 85%, and parcels from 52% to 95% (RBC Trends). The technical side of sorting is described in detail and without gloss by the PostTech team in its corporate Habr blog — from the protocols used to talk to machine controllers to the fate of envelopes with illegible handwriting. That publication is what makes this case verifiable: a rare instance of a state company exposing the internals of its production IT.

Problem

Manual sorting does not scale at these volumes: millions of envelopes and parcels a day with handwritten and printed addresses of varying quality. Each item must be read, its postcode and address determined and cross-checked, and the item routed correctly — within fractions of a second, or the logistics center's conveyor stalls. In numbers: the belt moves at 1.8 meters per second, and the routing decision must be made in about one second — before the first junction where a pusher either ejects the letter into the right chute or does not.

The main enemy of automatic recognition is real mail. Handwritten postcodes, faded ink, crumpled envelopes, addresses written over old ones, damaged barcodes. The classic dilemma for the system's designer: demand one-hundred-percent OCR confidence and the flow stops; let the machine 'guess' and letters ride off in wrong directions, with the cost of error measured in days of return shipping.

A separate engineering problem is the equipment zoo. Sorting machines from different manufacturers (Toshiba, Siemens, Vanderlande) speak different protocols, and without a unified software layer every center becomes an isolated island with its own logic. Finally, sorting is only half the task: even a perfectly sorted item travels along a route someone must build, and the monthly length of Russian Post's routes totals 54 million kilometers (CNews). Optimizing that network manually runs into the same limits of human capability as reading envelopes.

Solution

At the Post's logistics 'factories', letter sorting works like this. A letter on the belt passes under a camera that photographs the item in five projections — recognition quality is won already at image capture. OCR finds text in the shots, reads the postcode and address, and cross-checks them; the letter-sorting machine processes 40,000 letters per hour — the famous 11 letters per second (Vnukovo runs two such machines for 80,000 letters per hour, and the center as a whole passes 125,000 letters per hour). The parcel machine sorts up to 8,000 parcels per hour, and a robotic manipulator works parcel sorting at 4 seconds per item.

The key architectural decision is human-in-the-loop: if the text is unreadable, the item's photo automatically goes to video coding — an operator who types in the postcode manually from the image. The human backs up the automation instead of stalling the whole flow: the conveyor never stops over bad handwriting, and machine-unread letters rejoin the stream after manual entry. Problem items get a 'Reject' chute at the end of the belt, and operators read damaged barcodes with handheld scanners.

Underneath lies PostTech's own software platform — SMAB (Sorting Machine Automation Bridge), which ties machines from different manufacturers into one system: it talks to controllers over the industrial OPC UA protocol, events fly between devices via an MQTT broker, and routing decisions come from the 'Sortmaster' service (the machine literally asks: 'where do I put postcode 107037?'). The stack is Spring, PostgreSQL, Vert.x, Consul, React, Docker (Habr, PostTech blog).

The next loop is routing. In May 2024, the Post, together with the Skolkovo Foundation (VEB.RF group), began deploying an intelligent logistics modeling system on the Teraplan platform: AI analyzes historical route data, forecasts item inflows at each stage of the chain, and proposes optimal routes while cutting unnecessary transport nodes; the program spans 2024–2026. By April 2026 the system was live: it merges data on mail movement between sorting centers, builds routes for different transport types, and lets the company simulate network scenarios across up to 30 million operations daily (CNews, Online47).

The contact center runs a voice assistant on Yandex SpeechKit with 92–95% speech recognition accuracy (TAdviser) — a case of reusing off-the-shelf speech tech instead of in-house ASR development. There were shuttered experiments too: Yandex delivery-robot trials (2021–2023) were wound down over high equipment costs (Hi-Tech Mail.ru) — a detail that keeps the picture of the Post's technology portfolio realistic.

Result

The automated pipeline sustains a national-scale flow. Confirmed operational figures: 40,000 letters per hour on the letter-sorting machine (11 per second), up to 8,000 parcels per hour on the parcel machine, 3 million items a day — the capacity of the Vnukovo center alone, 4 seconds per parcel for the robotic manipulator, 5-projection photography for OCR, and 92–95% speech recognition accuracy for the voice assistant. Since 2026 a routing loop has joined the sorting loop: the Teraplan platform plans transport across a network of 1,100 sorting nodes and 38,000 post offices.

The routing loop's effects are so far described qualitatively, not quantitatively: reduced planning time, lower operating costs, and better forecast accuracy through clustering of logistics flows are claimed — no measured public figures for Teraplan exist yet, which honestly reflects the project's stage. What is absent from the public record entirely — and worth stating: the company has not disclosed the financial effect of sorting automation. The integral quality dynamics (on-time delivery: letters 54% → 85%, parcels 52% → 95% over 2013–2016) relate to the logistics modernization as a whole — new centers, transport, processes — not to OCR alone. The share of automatically recognized items and the cost of video coding are not published either. A separate frame is volume dynamics: the peak '8 million items a day' belongs to earlier overview materials, while 2024 publications already cite an annual letter flow of 1.3 billion.

In our view, the engineering essence of this case is not OCR as such (postal services worldwide have recognized addresses for decades) but an honest failure architecture: five projections at capture, the machine decides within a second, everything unreadable goes to a human video coder, the hopeless goes to reject. It is a conveyor designed around the assumption that AI will make mistakes — and therefore it never stops. The reverse order — 'first believe in 100% accuracy, then act surprised' — would have cost the Post stopped conveyors.

A second observation: the SMAB platform that tied Toshiba, Siemens, and Vanderlande machines together with a single protocol layer is, in our view, the case's most underrated part. Integration software never makes press releases, yet it is what turns a set of expensive imported machines into a manageable network — and it stays with the company when hardware suppliers change.

Lessons learned

1. Human-in-the-loop is a working architecture for OCR on messy data: unrecognized items go to a human video-coder, so bad handwriting never halts the conveyor.
2. Design around the assumption that AI errs: five projections at capture, one second to decide, manual entry and a reject chute for the rest — the system stays alive precisely thanks to these 'emergency exits'.
3. Throughput (40,000 letters/hour, 4 sec/parcel) are the honest sorting metrics; the company publishes no ruble savings, and the case must not invent them.
4. Five-projection photography shows OCR quality is won at image capture, not just in the recognition model.
5. The integration layer is worth more than individual machines: the SMAB platform (OPC UA + MQTT) tied Toshiba, Siemens, and Vanderlande equipment into one network and remains an asset when suppliers change.
6. Integral quality indicators (on-time delivery) grew over years of overall modernization — attributing the whole effect to one technology would be incorrect.
7. Off-the-shelf tech and shuttered experiments are part of a mature portfolio: SpeechKit covered the contact center without in-house ASR, and the delivery robots were honestly wound down when the economics failed.

Frequently asked questions