VISION ENGINEAvailable

See, track, and act on-device

A model registry with custom model upload, on-vehicle inference, and visual-inertial navigation. Detect generic classes like person and car, track them frame to frame, and feed the flight stack. No cloud calls.

View on GitHub Browse extensions

Available

Usable today. Early access — some pieces run behind a feature flag and APIs may shift between releases.

Accelerator backends

Cloud calls

VIO

GPS-denied nav

Open

Model registry

How it works

Camera to flight stack, all on the companion

Frames flow from the camera through inference and tracking, then out as guided setpoints. Nothing leaves the vehicle.

Vision pipeline

CameraCSI or USB

Frame capture

InferenceNPU / GPU / CPU

Object detectionperson, car, generic classes

TrackingTrack and project

Pixel to groundtrack id and lock

MAVLinkTo the flight stack

Guided setpoints

Detection runs on the board's NPU, GPU, or CPU. Tracks are projected from pixels to ground coordinates and handed to the flight controller over MAVLink.

Capabilities

A registry, not a single model

Model registry

Fetch, download, and cache detection models on the companion. Models stay on the device and run locally.

Custom model upload

Sideload your own model through the ground station. A checksum is verified before it is added to the catalog.

On-vehicle inference

The agent detects the board's accelerator and available compute, then picks the matching model variant automatically.

Vision navigation

Optical flow and monocular visual-inertial odometry hold position where GPS is unreliable or unavailable.

Hardware

Six accelerator backends

The registry carries model variants for each accelerator. Selection is automatic, based on the board and the compute it advertises in TOPS.

Rockchip RKNN0.8-6 TOPS

Qualcomm HexagonSNPE / QNN

NVIDIA TensorRTJetson, 40+ TOPS

Amlogic NPU5 TOPS

Google Edge TPUCoral, 4 TOPS

CPU / ONNXAny Linux board

Vision navigation

Fly where GPS cannot

The vision navigation extension installs through the extension registry and adds optical flow plus monocular VIO.

With vision navigation, a drone holds position and navigates where GPS is unreliable, indoors, under a canopy, or near tall structures. It supports ArduPilot 4.5+, PX4 1.14+, and iNav 7.0+, with an auto-detect probe for the supported camera and compute setups.

Explore the stack

Sensor FusionIn Development

Plug-and-play multi-sensor fusion hub on the companion computerLearn more →

AtlasAvailable

Capture-to-reconstruction world model: splats, point clouds, meshLearn more →

Put vision on your drone

The model registry and vision navigation are open-source and documented on GitHub. Bring your own model or start from the catalog.

View on GitHub