RFID Drones Over High-Bay Racking: Fly a Mission, Not a Photo Pass

A drone that flies a lap of the high-bay racking at night sounds like the last piece of warehouse automation. What it actually delivers is a set of observations, each tied to a position, a time and a flight — and whether that becomes inventory truth depends entirely on the model behind it. The flight log is the evidence; the stock figure is a conclusion drawn from it, and the two should never be confused when a discrepancy is argued about a month later.

What a drone can and cannot see

It is an upper-level instrument, not a replacement for a ground robot.

• It flies a fixed altitude band above the top beam and looks down at labels and tags on the upper levels.
• It cannot pick, count loose items inside a tote, or read a tag on the far side of a pallet.
• It sees the top face of cartons, which is often exactly where nobody ever placed a label.
• In narrow aisles the rotor wash moves shrink wrap and dust, and read rates move with them.
• Treat it as an upper-level tool and keep ground coverage for the lower two metres.

Flight is a mission, not a photo pass

A flight that is not recorded is not evidence.

• Define the mission: zone, rack rows, altitude band, speed, gimbal angle and planned waypoints.
• Record the map version and the coordinate frame the waypoints refer to.
• Bind every capture to a timestamp, a pose and the mission identifier.
• Log aborted legs and re-flights, because coverage gaps look exactly like missing stock.
• Keep the raw flight log even after the summaries have been imported.

The high-bay geometry problem

Height turns a simple read into a trigonometry exercise.

• At eight metres the read cone is wide; one pass covers several levels or none, depending on angle.
• Beams, wire mesh guards and pallet overhang all block line of sight to the label.
• Tags on the top beam are readable; tags on the pallet face usually are not.
• A skewed flight path pushes reads into the neighbouring bay and creates phantom variances.
• Resolve height ambiguity by combining the pose with the bay identifier read at floor level.

Read quality follows flight pattern

Most drone RFID failures are flight pattern failures.

• Fly at a consistent offset from the rack face; distance variation changes read rate more than tag type does.
• Slow down at corners, where the antenna sweeps past labels at a grazing angle.
• Fly each aisle in both directions when the first pass lands below the coverage target.
• Keep rotor wash away from loose film and packaging so two passes are comparable.
• Record signal strength per read, so weak observations can be separated from confident ones.

Safety and regulatory envelope

Indoor flight avoids airspace law, not workplace safety law.

• A written risk assessment is still required, and it should name the aisles, times and people involved.
• Define exclusion zones around staff, forklift lanes and charging areas.
• Set a battery reserve that guarantees a controlled landing rather than a best effort.
• Put the geofence in the flight controller, not only in the mission planning software.
• Log every manual override; overrides are the clearest signal that the mission design is unsafe.

From captures to a variance document

The ERP should never receive a raw drone read.

• Land captures in a staging table with mission, pose, timestamp, tag and signal strength.
• Build expected quantities from Odoo quants at mission start, then compare location by location.
• Classify each difference: real shortage, misplacement, unreadable tag, blocked shelf, stale map.
• Attach a confidence value to every line so reviewers can sort by reliability.
• Publish the comparison before proposing stock moves, and keep the snapshot it was based on.

The sync model: missions, poses, snapshots

The integration is a data model question disguised as an API question.

• Model mission, leg, waypoint and capture as separate records with stable keys.
• Version maps and reference the version inside every capture.
• Store the expected snapshot used for comparison, together with its generation time.
• Allow late arrival; a re-uploaded mission must be idempotent and produce no duplicate tasks.
• Expose variance documents to Odoo as tasks with owners, never as silent adjustments.

Reading tags at height, on metal and in cold

The upper rack is an unfriendly radio environment.

• Metal racking reflects and cancels signal, and tag orientation matters more than tag generation.
• Freezer zones fog lenses and shift antenna impedance; test at temperature, not on the bench.
• On-metal tags cost more, and a mixed estate reads unevenly for reasons that are hard to see.
• Label condition degrades with forklift contact; unreadable tags cluster at the impact points.
• Track read rate per bay over weeks, so a failing tag surfaces before it becomes a variance.

Where drones beat ground robots, and where they do not

Choose the platform by the geometry, not by the demonstration.

• Drones win above four metres, in aisles too narrow for a mast, and where the floor is inaccessible.
• Ground robots win on endurance, payload and reading at eye level under stable lighting.
• Drones are weaker in draughts from open doors, in dust, and where ceiling height is limited.
• A mixed fleet should still feed one variance model, not two parallel reporting pipelines.
• Compare cost per thousand locations observed, not cost per unit.

Pilot design, acceptance and metrics

A pilot has one job: to show whether coverage improves on today's process.

• Start with one high-bay zone and a single category with clean tag data.
• Run in report-only mode for four to six weeks before proposing any adjustment.
• Measure coverage, agreement rate, false positives and time to close a variance.
• Watch the share of variances traced to map or tag problems; it should fall month by month.
• Define in advance what result would justify stopping, and honour that definition.