Model downloads

Models, built for the edge.

Visual positioning for GNSS-denied navigation, packaged to run on the hardware a real vehicle carries. Every release ships as one versioned, checksummed bundle, and states on its own entry exactly which formats and which evidence it contains — because not every release contains the same things, and we would rather say so than let a page make a promise a particular model cannot keep.

Visual positioning for GNSS-denied navigation
Runs on ARM and x86; edge accelerator builds where stated
Every release states the formats and evidence it ships
Versioned, checksummed, documented

What a release ships.

One directory per release, always including a model card describing intended use, measured performance and known limits, and a SHA-256 checksum for every file. Beyond that the contents vary by model and are listed on each entry: ONNX exports for CPU inference on ARM and x86, the training checkpoint, compiled accelerator builds, and verification reports where they exist.

Deployment parity is measured, not assumed: each exported build is verified against the reference implementation on real flight imagery, and the report ships in the bundle.
Where a quantised or accelerator build is offered, the accuracy evidence comes with it: the INT8-vs-fp32 comparison is published alongside, so you can see the cost of the speed-up before you deploy it. A release without that evidence does not offer the build.
Full provenance per release: a semantic version, the exact source checkpoint, per-file checksums, and a model card covering evaluation scope and known limits.

The models.

In qualification — evaluation builds now

Visual navigation model

Our visual-positioning model for when satellite navigation is jammed, spoofed, or unavailable. It fixes position by referencing live camera frames against imagery of the operating area, on the vehicle, offline. Qualification is under way; evaluation builds go to partners now, and the model card travels with them.

Version: semantic version, assigned on release
Formats: PyTorch checkpoint (.pt) · ONNX · Hailo-8 HEF
Target hardware: Hailo-8 NPU · ARM64 · x86-64 CPU
Model card: included in the bundle, and available on request before purchase.

Request access

Evaluation release — measured, limited envelope

Badb Vehicle Detector

Cues an operator to vehicles in aerial video, on the vehicle or in the cloud. Trained from scratch on a fully cleared licence lineage — no ImageNet, no COCO, no copyleft — so it can ship inside a closed defence product. Measured where detection is hardest: on real combat footage from Ukraine, much of it carried over degraded analogue radio links, it cues roughly six vehicles in ten at fewer than one false alarm per three frames, rising to about three in four where a higher false-alarm rate is acceptable. Recall improves with target size — small and distant is the hard case. Validation to date is on that single domain, and the model card states what is proven, what is assumed and what is untested.

Sizes: a small model for edge hardware, a larger one for workstation and server
Formats: ONNX (this release ships ONNX only — no edge-accelerator build yet)
Target hardware: ARM64 · x86-64 CPU · hosted API
Model card: included in the bundle, covering the conditions it is valid in and the ones it is not

Try it in your browser Request the bundle

Evaluate it

Request an evaluation build.

Tell us your platform, runtime, and operating area, and we'll deliver the appropriate release as a time-limited download link — no account needed. Builds tuned to your camera, altitude band, and terrain are available under contract.

Request access