See. Track. Act. All on-device.
31+ autonomous vision behaviors running on NPU-accelerated companion computers. Object detection, tracking, obstacle avoidance, follow-me, precision landing, crop analysis, and more. Zero cloud dependency. All open-source.
Architecture
Layered architecture
Core Engine
Frame source, NPU detector, KCF tracker, pixel-to-ground projector, model manager. The foundation that every behavior builds on.
Obstacle Avoidance
Multi-sensor depth fusion, 3D occupancy grid, BRAKE and DETOUR planning, APAS for manual flight. Always-on, injected into every behavior's update cycle.
Behavior Layer
31+ exclusive behaviors configured by the active suite. Only one flight behavior runs at a time. Safety behaviors (obstacle avoidance, terrain following) run in parallel.
Behaviors
31+ behaviors across every vertical
Consumer
Follow-me, orbit, spotlight, quickshots (dronie, rocket, circle, helix, boomerang), precision landing, gesture recognition, panorama, hyperlapse, mastershots, hand launch and landing.
Agriculture
Crop row following, weed detection and targeted spray, plant health mapping, terrain following.
Inspection
Defect detection (crack and corrosion), thermal anomaly classification, structure following (power lines, towers).
Search & Rescue
Person detection (RGB and thermal), search pattern automation, night vision enhanced, wildlife tracking.
Safety
Obstacle avoidance (multi-sensor, APAS), terrain following (vision-based AGL), landing zone assessment, perimeter monitoring.
Survey
Camera trigger, image quality monitor, coverage tracker, GSD calculator.
Hardware
Runs on what you have
Rockchip RKNN
0.8-6 TOPS. RK3566, RK3588, RV1126. Entry to mid-range companion computers. 8-30fps depending on model size.
Qualcomm Hexagon
Snapdragon NPU via SNPE/QNN runtime. Common on mobile and embedded AI platforms. 4-26 TOPS.
NVIDIA GPU / Jetson
40+ TOPS via TensorRT. Enterprise-grade. 30fps at full resolution with multiple concurrent models.
Amlogic NPU
A311D, S905D3 with integrated NPU. Cost-effective boards with 5 TOPS performance.
Google Edge TPU
Coral accelerator via USB or M.2. 4 TOPS int8 at 2W. Ideal for low-power deployments.
CPU Fallback
Any Linux board via ONNX Runtime. Lower framerate but fully functional for development and testing.
Safety
Obstacle avoidance that actually works
Multi-sensor depth fusion: monocular depth estimation, stereo cameras, LiDAR, ultrasonic sensors. 3D voxel occupancy grid covering 40x40x20 meters at 0.5m resolution with 5-second decay.
Velocity-based stopping distance calculation with RED, YELLOW, and GREEN zones. Two avoidance modes: BRAKE (stop and hover) and DETOUR (path around obstacle). APAS for manual flight blends pilot stick input with avoidance corrections.
Obstacle-aware return-to-home with A* path planning.
Explore the full behavior catalog
Every behavior is open-source and documented on GitHub.
View on GitHub