Autonomous
Inventory
Robot

Warehouse Robotics / Multimodal Sensing / Point-cloud ProcessingProject Materials / 2024–2026

Overview

An autonomous warehouse inspection robot that coordinates mobile navigation, multi-height sensing and inventory reconciliation across stacked storage.

My work focused on connecting the robot’s sensing and control stack: coordinating AGV movement, PLC-driven camera positioning, RGB and depth capture, RFID collection and the operator-facing processing interface.

ScopeSystem integrationSensor integrationPython developmentPoint-cloud processing

One inspection run / Four coordinated systems

MOVE. POSITION.
CAPTURE. RECONCILE.

01Navigate

The AGV follows the warehouse route and pauses at defined inspection points.

02Position

A PLC moves the sensing carriage through lower, middle and upper capture heights.

03Capture

Left and right RGB-depth pairs record shelf images and point clouds while RFID captures tag data.

04Reconcile

The interface combines barcode results, RFID records and estimated box counts for review.

Physical system / Multi-height inspection

EXTEND THE ROBOT’S VIEW
FROM FLOOR TO HIGH SHELF.

A vertical sensing mast turns a mobile base into a two-sided inspection platform. At each stop, the carriage moves through three heights so one route can cover both sides of a multi-level warehouse aisle.

Autonomous inventory robot positioned between warehouse racks — FULL PLATFORM / AGV + VERTICAL SENSING MAST

Close view of the RGB and depth camera pair mounted on the sensing mast — RGB + DEPTH CAMERA PAIR

RFID antenna mounted on the vertical sensing carriage — RFID ANTENNA / ITEM IDENTIFICATION

Close view of the AGV base with onboard computers, power electronics and cabling — ON-BOARD COMPUTING + PLC INTEGRATION

Algorithm implementation / Point-cloud counting

FROM DEPTH POINTS
TO A BOX COUNT.

Each capture is indexed by inspection point, aisle side and carriage height. The Python pipeline pairs the RGB image with its PLY or TXT point cloud, filters the measured scene and estimates quantity from occupied volume.

Index the capture

Organise left and right observations at lower, middle and upper heights, then pair each image with its corresponding point cloud.

POINT / SIDE / HEIGHT

Filter the scene

Apply a 0.01 voxel size, crop to a calibrated 3D region and remove statistical outliers before measurement.

5 NEIGHBOURS / STD RATIO 2.0

Integrate volume

Use a 10-neighbour search in the XZ plane to estimate local point spacing, then accumulate local area multiplied by depth from the reference plane.

V ≈ Σ dᵢ² · (yref − yᵢ)

Count and review

Divide the integrated volume by the known box volume, round the estimate and review it beside barcode and RFID records in the operator interface.

N = ROUND(V / (500 × 310 × 360 MM³))

Control and perception pipeline

SENSE. FILTER. COUNT.

The Python control layer connects AGV commands with camera and RFID acquisition. The processing layer filters each point cloud to the shelf region, estimates occupied volume and compares the count with item-level identification data.

Architecture diagram linking AGV control, image and point-cloud acquisition, barcode recognition and stacked-box counting — SYSTEM ARCHITECTURE

Route

Send the AGV along its inspection path and monitor arrival at each capture point.

Acquire

Trigger lower, middle and upper RGB-depth capture on both sides of the aisle.

Process

Filter each point cloud to the shelf region and estimate stacked-box quantity.

Review

Compare expected and observed counts with barcode and RFID evidence in the interface.

Outcome and validation

Sense. Identify. Count.

One workflow combines three-height RGB-D capture, RFID/barcode identification and volume-based box counting. Preliminary warehouse tests are documented; customer-site validation remains open.