Artificial Intelligence and RFID Robotics in the Recreational Vehicle Industry: Results of the RUE 47 Project

Published by CMC MOTO CLASS COMPANY S.R.L. (ATVRom) — 21.08.2026

We have finalised the experimental development phase of the project “Adoption and Development of Advanced Digital Technologies at CMC MOTO CLASS COMPANY S.R.L.”, funded through the National Recovery and Resilience Plan. The project produced two digital components: an AI-based recommendation engine, with a conversational assistant integrated into the online store, and an automated RFID inventory system, built around an autonomous mobile robot.

We present below the main results, including the design baseline established during the conceptual phase and the solutions that were ultimately discarded - because in a research project these have the same value as the successes.

Project details

Funding contract: No. 47.1.i3.c9 of 29.04.2025

Project RUE: 47

Partnership leader: CMC MOTO CLASS COMPANY S.R.L. (commercial brand ATVROM), CUI 36145193

Partner: GLOSOFT SOLUTIONS SRL, CUI 38032549

Total project value: 11,231,818.70 lei

European Union contribution: 6,635,610.66 lei (non-repayable grant funding from the PNRR – NextGenerationEU)

Implementation period: 29.04.2025-28.08.2026

Implementation location: Micesti Village, Micesti Commune, National Road 73 Pitesti–Campulung, No. 74M, Arges County

The problem: why standard solutions do not work

The recreational vehicle trade combines two incompatible logics. The vehicle is an individual asset, tracked by serial number or VIN, expensive, with a long sales cycle. Parts and accessories are fungible goods, inexpensive, with fast turnover. A system that treats them identically fails in both directions.

When it comes to recommendations, commercial engines optimise a probability: what a similar customer would buy. In our field, however, there is a prior, purely physical question: does the part fit the vehicle? An incompatible recommendation is not a weak recommendation — it is a return, a shipping cost and a loss of trust. Compatibility cannot be one signal among many; it must be an eliminatory filter, applied before any statistical ranking.

When it comes to inventory, RFID technology is commonly presented as a solution for instant stocktaking, but it carries a fundamental logical asymmetry: a tag that is read proves the item is present, while a tag that is not read does not prove the item is absent. An item may remain unread because of metal shielding, antenna orientation, its position on the shelf or a damaged tag. A system that automatically turns inventory discrepancies into stock adjustments will write off items that physically exist.

Result I - the AI Optimisation Engine

The component is organised in three layers, each narrowing the search space of the next: a deterministic, rule-based compatibility layer (eliminatory - a product that does not pass it appears in no recommendation); a statistical ranking layer, which combines historical cross-selling, semantic similarity and in-stock availability; and a serving layer towards the online store.

Volumes at the time of measurement: 1,030 vehicles with pre-computed recommendations, 17,070 products with tracked availability, over 1.8 million rows in the training set.

Measured results (production data, 17.08.2026)

Serving latency (median): 51 ms — design baseline < 500 ms

Latency (95th percentile): 77 ms — design baseline < 500 ms

Recall@50: 0.5010, target ≥ 0.40 met — heuristic baseline 0.4755

NDCG@10: 0.3210, target ≥ 0.35 not met — heuristic baseline 0.2923

Hit rate in the first 12: 0.7647 — heuristic baseline 0.7549

Area under the curve (destination classifier): 0.9934 / 0.9969, target ≥ 0.70 met

The trained model outperforms the heuristic baseline on all ranking metrics, but on NDCG@10 it remains below the target set during the conceptual phase. We report both figures without restating the target: it was set before the real data was known and proved optimistic for the signal density available.

The conversational assistant: the prohibition on knowing

The assistant in the online store works on a principle that runs counter to current practice: the language model is forbidden from using its own general knowledge. It answers exclusively from the data returned by the tools that query the recommendation engine, and for questions outside that data it hands the conversation over to a member of the team.

The rationale is a cost asymmetry: a plausible but wrong specification about a specific vehicle is more harmful to the customer than an answer that acknowledges its limits. A sales assistant is not judged by how many questions it can cover, but by how rarely it misleads.

The component has been classified as a limited-risk system under Regulation (EU) 2024/1689, with the associated transparency obligations: labelling automatically generated content, explaining the reason behind a recommendation and offering the option to opt out of personalisation.

Result II — the Advanced RFID System

The system is installed and operational at the Micesti warehouse and comprises three subsystems: a tag encoding station (a printer with an integrated radio module, European 865–868 MHz band), an autonomous mobile robot carrying a reading terminal, and a handheld terminal for spot checks.

The tags use a proprietary 96-bit encoding scheme developed within the project. A tag names an item, not a commercial reference - renaming a product in the catalogue does not invalidate a tag that has already been printed. Decoding happens on the server, not in the reader, because a terminal holding an outdated copy of the catalogue could miscount stock.

The central rule: the human approves, the system prepares

The most important design decision is counterintuitive in an automation project: no stock adjustment reaches the ERP system without the explicit, line-by-line approval of a human operator. Automatic approval thresholds have been removed entirely - in both directions, for surpluses as well as shortfalls.

The reason follows directly from the asymmetry described above: an automatic stock decrease would be a decision taken on the basis of missing evidence. The useful automation here is not the decision, but the complete preparation of the decision - the physical pass, the reading, the decoding, the reconciliation with the book records and the presentation of discrepancies with confidence indicators attached. The time saved comes from removing manual counting, not from removing the operator.

The system additionally reports the reading convergence curve: if the last pass of the robot was still discovering new tags, every stock decrease is forced into human review.

What did not work

SLAM navigation was implemented and abandoned. The warehouse aisles are long and lack distinctive geometric features, and localisation by map matching accumulated drift - a problem of the environment, not of the implementation. The solution adopted is following a physical guide line: less general, but deterministic. For a night-time stocktake along a route that repeats identically, the fixed route is not a limitation; it eliminates an entire class of errors.

The catalogue coverage target cannot be reached with the current data. Compatibility is known for 78.1% of equipment, but for only 13.8% of accessories - 46.3% in total, against the 60% target. The cause has been identified: for accessories, compatibility information is encoded implicitly in the part number of the manufacturer, in a convention that differs from one manufacturer to another and cannot be extracted systematically. The ceiling is a property of the catalogue, not of the algorithm - no algorithmic method can overcome it. The conclusion redirects the effort from model optimisation towards completing the catalogue data.

Current status and next steps

The recommendation engine and the conversational assistant are implemented and measured on production data. Access was deliberately restricted to the project team for the duration of testing, as a quality measure — which is why the figures above are technical system measurements, and the article makes no claim regarding the effect on customer behaviour: no such effect has been measured yet.

The RFID system is built, installed and operational, validated through simulation across eleven scenarios before being put on hardware. The actual transmission of corrections to the ERP remains disabled until the field contract with the ERP system vendor is validated; until then the flow works in full, up to and including approval.

Directions for continuation: the first complete stocktakes on real hardware, with the read rate measured by item category; opening public access to the recommendation component and starting to measure business indicators from an established baseline; completing the compatibility data for accessories; extending the robot to closed-loop control.

This publication disseminates the results of Activity 5 (Experimental development, Article 25 of EU Regulation No. 651/2014) within Activity 7 — Transfer of research and development skills, competences and knowledge.