Robotics / 2023 / Validated in simulation

Autonomous Retail Intelligence Robot

A ROS-based autonomous retail robot combining real-time shelf perception, multi-object tracking, autonomous navigation, customer-request handling, and ERP synchronisation.

What this system solves

Brick-and-mortar retail still depends on manual stocktakes, paper customer requests, and inventory data that drifts out of sync with what is actually on the shelves. A robotic system that can see, navigate, and update business records continuously is operationally valuable but rarely shown end-to-end.

System overview

Built on the Husarion ROSbot 2 Pro platform in Gazebo, ARIR uses YOLO-family detectors with DeepSORT tracking to maintain a stable inventory view across frames, publishes product locations through ROS topics, synchronises state to an Odoo ERP via XML-RPC, and accepts customer requests that drive guided navigation through move_base.

Architecture

• RGB/Depth camera frames are consumed by a perception node running YOLO inference.
• DeepSORT maintains track identities across frames and feeds a debounce buffer that promotes a detection into inventory only after stable observation.
• An inventory aggregator publishes product locations on a ROS topic and persists changes to CSV and Odoo via XML-RPC.
• Independently, gmapping builds the store map while move_base handles path planning and obstacle avoidance.
• A customer-request handler converts requests into goal poses and routes the robot to the appropriate shelf coordinates.

Engineering decisions

What I built

• Designed the perception → tracking → inventory state → ERP pipeline as a single end-to-end system.
• Trained and integrated YOLO detectors for shelf product classes; tuned DeepSORT for stable IDs under motion.
• Built the ROS nodes that convert detections into product locations and publish them to the navigation and ERP layers.
• Implemented Odoo XML-RPC synchronisation with idempotent inventory updates.
• Set up the gmapping + move_base navigation stack and integrated customer-request handling.

Validation and results

• Autonomous navigation, perception, and ERP synchronisation evaluated in Gazebo with multiple shelf layouts.
• Customer-request flow validated end-to-end: voice/text input → product lookup → guided navigation → arrival.
• Detector evaluated against held-out images; tracker stability checked across multiple simulated camera passes.

Limitations

• Evaluation was primarily in simulation; physical robot trials were limited by hardware access.
• Dataset diversity was constrained — generalisation to unseen store layouts requires additional training data.
• Model training was cloud-dependent; the pipeline assumes pre-trained weights at runtime.
• Perception → ERP synchronisation introduces non-zero latency that needs further measurement on production-class hardware.

What would come next

• Hardware-in-the-loop evaluation on a real ROSbot 2 Pro with on-board inference.
• Switch from lexical product matching to embedding-based queries for fuzzier customer requests.
• Add health and freshness monitoring on the Odoo synchronisation path for production deployments.

Media gallery

Next