Industrial robotics · Tokyo
Japanese industrial: 80 AGVs, ROS 2, +40% warehouse throughput.
Slow fleet, collisions, downtime. Nav2 refactor, perception fusion, multi-agent planning. Zero collisions in 6 months.
KPI
+40%
warehouse throughput
Duration
14 mois
Team
7 engineers
Hub(s)
Tokyo
A fleet of 80 AGVs in a 42,000 m2 warehouse, two collisions per week and a capped throughput. The problem wasn't the robots, it was coordination.
The context
Japanese automated logistics industrial, 3,200 employees, Tokyo hub. Pilot warehouse near Nagoya, 80 AGVs deployed, aging ROS 1 navigation, monocular perception, individual planning with no fleet coordination.
The problem
- Warehouse throughput stagnant for 14 months
- Two collisions/week, one major every two months
- Cumulative downtime: 18% of production time
- Slow navigation in narrow corridors (60 cm margin)
- ROS 1 EOL scheduled, no more security support
The approach
Complete overhaul on ROS 2 Humble with Nav2, LiDAR + RGBD camera sensor fusion for 3D perception, multi-agent planning via Conflict-Based Search. Intensive validation in Isaac simulation before any physical deployment.
The technical pillars
- Perception stack in Rust (p99 latency under 30 ms)
- Customized Nav2 with dynamic costmap shared between agents
- CBS solver for real-time multi-AGV conflict resolution
- Isaac Sim simulation with digital twin of the full warehouse
- Rollout in batches of 10 AGVs, ROS 1 fallback maintained for two months
The stack
- ROS 2 Humble, Nav2, MoveIt 2
- Rust for perception and low-level control
- Cyclone DDS for middleware, fine QoS tuning
- Isaac Sim for validation, Foxglove for observability
- Intel RealSense D455 cameras, Velodyne VLP-16 LiDAR
The results
- Warehouse throughput: +40%
- Collisions: zero over 6 months post full deployment
- Cumulative downtime: 18% to 4%
- Average AGV speed in corridor: +28%
- Migration with no production interruption (24/7 maintained)
« The Abbeal team understood our Genchi Genbutsu culture: go see on the ground. They spent three weeks observing our operators before the first line of code. »
What we learned
Cyclone DDS is superior to Fast DDS for our case (deterministic latency), but requires serious QoS tuning. The CBS solver scales up to 80 agents, beyond that you need a hierarchical approach. Honest mistake: we underestimated LiDAR calibration time (two weeks more than planned). To redo: involve warehouse operators from the simulation phase, their feedback avoided three bad design decisions.
// Read next
Luxury jewellery & watchmaking · Genève + Paris + Tokyo
Cartier: from audit to in-house private LLM.
Compass (front + back architecture audits), Mapper (watchmaking + jewellery product generator), competitive data ETL on BigQuery, and now a private LLM fine-tuned on Cartier's own infra. A long-term tech partnership on the data and AI stack of a luxury house.
LLM privé
fine-tuned on Cartier infra
Tier-1 bank · Paris
BNP Paribas: Reference Book PO, from React/Redux to product AI agents.
Three Abbeal engineers at the core of the PO Marketplace. React/Redux/Node platform initially, now augmented with a product RAG, Claude agents for PM assistance, and an event-driven Kafka layer to scale.
RAG
PO product catalog
Digital banking / FinTech · Tokyo (Tamachi)
Money Forward: data backbone of a brand-new digital bank in Tokyo.
Money Forward, a Japanese FinTech leader listed in Tokyo, partnered with a top-tier Japanese banking group to launch a brand-new digital bank built from scratch. Abbeal partners on the Data Engineering side: designing and operating the Data Hub (Databricks + Delta Lake + dbt + AWS Tokyo) serving JFSA reporting, AML, risk management.
Data Hub
from-scratch digital bank Tokyo