Fusing UWB and VIO: Sub-10cm Forklift Tracking on Sparse Anchors

Why doesn't UWB alone hold sub-10cm on a moving forklift?

An 802.15.4z HRP UWB channel is 499.2 MHz wide, which resolves time of flight to roughly 10 to 30 cm line of sight. The problem is not the ranging, it is everything around it. A TDoA or multilateration fix needs three to four anchors with simultaneous line of sight, and a warehouse is built to block line of sight: 10 m steel racking, pallets of dense product, trucks in motion. A range measured through a loaded pallet arrives late and reads 30 cm to over 1 m long. Aisle geometry makes it worse, because the anchors visible from inside an aisle tend to sit along the aisle axis, so cross-aisle dilution of precision blows up exactly where placement accuracy matters. Then there is motion: at 10 Hz ranging and a forklift moving 3 m/s, consecutive fixes land 30 cm apart, and interpolation smears through every turn. The standard answer is anchor density, one every 15 to 20 m, each needing power, backhaul, and a scissor lift to mount. Density is a workaround, not a design.

What does visual-inertial odometry contribute, and where does it fail?

Visual-inertial odometry pairs a camera with a 100 to 200 Hz IMU and integrates both into a smooth 6-DoF pose. Pipelines in the ORB-SLAM3 and VINS-Fusion class drift around 0.5 to 1 percent of distance traveled in feature-rich scenes. That error profile is the opposite of UWB: over short windows VIO is centimeter-consistent and updates fast enough to track a turning forklift frame by frame. But it is a relative sensor. It does not know where the map origin is, and drift compounds silently: a 100 m run down an aisle accrues 0.5 to 1 m of error with no internal way to notice. Warehouses attack VIO's weak points directly:

• Long runs of visually identical racking invite wrong loop closures.
• Freezer sections run dark and fog lenses.
• Mast vibration at full lift corrupts the IMU and blurs frames.

Bare VIO keeps reporting confident poses while sliding off truth, which is worse than failing loudly. Neither sensor alone survives a full shift at sub-10cm. VIO is smooth, but it lies slowly.

How does tightly coupled fusion let one anchor do the work of four?

The architecture choice that matters most is tight coupling. A loosely coupled design fuses finished UWB position fixes, which means you still need three to four anchors in view to produce a fix at all, so sparse anchors buy you nothing. A tightly coupled filter, an error-state EKF or a sliding-window factor graph, consumes each raw two-way range as a single scalar measurement. With VIO constraining the shape of the trajectory, a handful of ranges to one known anchor, taken over a few meters of motion, pins both position and heading. Range-only SLAM uses the same trick. Three details decide whether it works in practice:

• Time sync. Hardware-timestamp UWB, camera, and IMU to well under a millisecond; at 3 m/s, 10 ms of skew is 3 cm of error.
• Lever arms. Calibrate tag antenna, IMU, and camera offsets to about 1 cm, or that error lands directly in the output.
• NLOS rejection. Gate each range by filter innovation and by first-path power versus total received power, and downweight late arrivals.

One honest range beats four noisy fixes.

How sparse can anchors actually get in a racked warehouse?

Standalone UWB TDoA deployments in racked warehouses typically land near one anchor per 400 to 800 square meters to keep three anchors in view everywhere. With fusion, the question changes from coverage to drift budget. At 0.5 percent VIO drift, a 40 m gap between ranging contacts accrues about 20 cm, and the next few ranges pull the estimate back. So you place anchors where two conditions meet: the forklift passes often, and accuracy pays. That means aisle ends, cross-aisle intersections, dock doors, and staging and pick zones, not a uniform grid. Inside that layout, hold sub-10cm where ranging contact is fresh, at the pick face and the dock, and accept 20 to 30 cm mid-aisle where nothing is being placed. Most forklift use cases, pallet put-away verification, dock door assignment, near-miss analysis, only need the tight number at task locations. The honest spec for your ops team is accuracy per zone, not one number for the whole building. Put anchors where accuracy earns money.

What breaks in production, and how do you harden it?

The filter is rarely what fails in production. The forks are. You care about pallet position, and the forks travel up to 10 m vertically on a mast that flexes under load, so either instrument the mast with an encoder or track the fork carriage visually and model the offset explicitly. Inventory changes the radio environment week to week: a rack full of canned goods blocks paths that were clean when it stood empty, so NLOS handling must be per-measurement, not a calibration map from install day. Cameras need maintenance reality: dust film, condensation at freezer doors, lens heaters where needed. And the quiet killer is anchor survey error. If anchors are placed into the map with 5 cm of error, no filter recovers it. Survey with a total station or run a SLAM-based auto-survey, then re-verify after any racking change. Your survey is your accuracy ceiling.

Should you build this fusion stack or license it?

A production-grade estimator of this kind is a 12 to 24 month project for a team spanning state estimation, UWB firmware timestamping, and calibration tooling, and those engineers are among the scarcest hires in robotics. The second cost is quieter: fusing RF ranging with inertial and optical tracking is dense patent territory, and discovering that after launch is expensive. Position Imaging licenses a portfolio of granted US patents in real-time positioning, radio-frequency ranging, computer vision, and machine learning, cited by Apple, Bosch, and other major firms, including US 11,774,249, US 12,079,006, US 12,066,561, and US 12,000,947. Licensing gets you two things at once: proven methods you do not have to reinvent, and freedom to operate in the exact fusion techniques this post describes. Teams that license spend their engineers on the product, not on rebuilding the Kalman filter, and ship tracking features in months instead of years. License the math and ship the forklift.