PAIRS Lab

RIVF 2026 International Conference on Computing and Communication Technologies

Rack-to-Bin

Hierarchical localization and visual refinement for autonomous UAV warehouse inspection

Van Tong Nguyen, Thang Tran Viet, Van Bang Nguyen, Xiem HoangVan*, Thanh Nguyen Canh

VNU University of Engineering and Technology   (T. N. Canh also with JAIST)   *Corresponding author

Code Video
A UAV inspects storage bins in a narrow-aisle simulated warehouse.

01 · Abstract

One rack tag, every bin, without GPS

Autonomous UAVs offer a flexible way to inspect inventory in indoor warehouses, where GPS is often unavailable and storage racks stand in narrow aisles. This paper presents an integrated UAV inspection system built on a fiducial-anchored coarse-to-fine strategy.

The UAV first navigates toward the target rack using a LiDAR-based OctoMap and a collision-aware path planner. At the rack, a rack-level AprilTag is detected and combined with the known rack geometry to estimate the nominal position of the selected bin. Because small position errors can remain after global navigation, a proportional visual-servoing step uses the QR code's position in the camera image to refine the final inspection pose before decoding.

In a custom Gazebo warehouse with narrow aisles, under Point-LIO localization, the method raises the capture success rate from 10.42% to 86.05% of eligible trials and reduces the mean per-axis capture error to about 2.3 cm. Local visual refinement further reduces the image-space alignment error from 212.67 to 24.34 pixels.

Inspection mission in simulation: rack approach, AprilTag detection, target-bin approach and QR code inspection.
Fig. 6: an inspection mission in simulation, from rack approach to QR code inspection.

02 · Video

Watch the inspection mission

03 · Method

Coarse to fine: rack, then bin, then image

System architecture: an operator and inventory backend sends target rack and bin commands to the inspection mission layer (mission manager, target localization and approach, local inspection), which sends navigation goals to the UAV autonomy stack (mapping, path planning, control) running on the hardware (RGB-D camera, 360 LiDAR, PX4).
Fig. 1: overall architecture of the UAV warehouse inspection system.
  1. 1

    Collision-aware navigation

    A LiDAR OctoMap and an A* planner take the UAV through narrow aisles to the target rack, with a reactive bumper as a safety layer.

  2. 2

    Rack-level anchor

    One AprilTag per rack, plus the known rack geometry, gives the nominal inspection pose of every bin on that rack.

  3. 3

    Boustrophedon sequence

    Bins on a rack are visited in a snake-like (boustrophedon) order, column by column along each level.

  4. 4

    Visual servoing on the QR code

    A proportional controller centres the bin's QR code in the camera image, absorbing the error left by global localization.

  5. 5

    Decode and verify

    The decoded QR payload is checked against the inventory database: matched, mismatched or unread.

  6. 6

    Real-platform design

    The simulated UAV mirrors a Pixhawk 6X + PX4 drone with a Jetson Orin NX, Livox Mid-360 LiDAR and RealSense D435i camera.

OctoMap 3D occupancy of the warehouse and an A* path planned through it.
Fig. 3: OctoMap occupancy and the planned A* path.
QR code alignment error in the image plane: horizontal and vertical offsets of the QR center from the image center.
Fig. 4: QR alignment error in the image plane.
Boustrophedon inspection order over a rack with three columns and six levels.
Fig. 2: boustrophedon inspection order on one rack.
Simulation setup: the Gazebo warehouse with narrow aisles and the simulated UAV with Livox Mid-360 and RealSense D435i.
Fig. 5: simulation environment and simulated UAV.

04 · E1 · Capture accuracy

How close does the UAV get to each bin?

54 requested bins per variant. A capture succeeds when the horizontal and vertical errors are both under 5 cm. Pick the localization source and a metric.

Servoing roughly doubles the time per bin (about 9–10 s instead of 4–7 s), and some bins are never reached: success over all 54 requested bins is 76% (GNSS) and 69% (Point-LIO).

05 · E2 · Narrow aisles

Safety versus reaching the goal

30 paired trials per configuration. Collision-free counts trials without contact; success also requires reaching the goal within 0.60 m in 20 s.

Collision-free trials% · higher is better
Goal success% · higher is better

The planner makes flight far safer (96.7–100% collision-free) but does not raise goal success, and a large bumper distance (1.0 m) causes conflicting corrections in tight aisles. The A* planner and threshold bumper are deliberately simple and can be swapped for stronger planners.

06 · E3 · QR inspection

Centring the QR code before reading it

Image-space alignment errorpixels · lower is better
Inspection time per binseconds · lower is better

Local alignment cuts the alignment error from 212.67 to 24.34 px, but the nominal pose already decodes 100% of codes, so the gain is viewpoint accuracy rather than decoding rate, at about 11 s per bin.

Inspection demo from the project repository (MuJoCo simulation).

07 · Cite

Citation

@inproceedings{nguyen2026racktobin,
  title     = {Rack-to-Bin Hierarchical Localization and Visual Refinement for Autonomous UAV Warehouse Inspection},
  author    = {Nguyen, Van Tong and Tran Viet, Thang and Nguyen, Van Bang and HoangVan, Xiem and Nguyen Canh, Thanh},
  booktitle = {RIVF International Conference on Computing and Communication Technologies (RIVF)},
  year      = {2026}
}