See everything.
Train on anything.
RAPTOR is a real-time AI detection, classification, and targeting system that runs on a $250 edge computer. Operators correct the AI in the field and retrain custom models—no cloud, no PhD, no datacenter.
Real-time AI-Powered Recognition, Analysis & Persistent Tracking for Observation on Robots
$250
Compute module — NVIDIA Jetson Orin Nano
30+ FPS
Detection with TensorRT; 15+ FPS with full tracking
< 30 ms
Inference latency — low enough for terminal guidance
60 g
Module weight — fits 5″ FPV frames
The feed should be watched by software, not by a tired operator
Today
An operator stares at a video feed for hours so nothing is missed. A jammed link means a blind aircraft. And when a new threat shows up in theater, teaching the system to recognize it means a vendor contract and a wait measured in months.
With RAPTOR
The model watches the feed and alerts on detections. Visual target lock rides through RF jamming and signal loss. And when the AI gets something wrong, the operator corrects it on the spot and retrains on-device—the system knows the new threat by the next mission.
The only field-retrainable AI vision system at the edge
Enterprise competitors cost 10–100× more, lock you into proprietary hardware, and can't learn new threats in the field. RAPTOR can.
Train in the field
Operator corrects a misdetection. RAPTOR learns. Retrain custom models on-device without sending data anywhere. The AI gets smarter every mission.
$250 compute module
Runs on NVIDIA Jetson Orin Nano. Full GPU-accelerated inference at a fraction of enterprise pricing. Shield AI charges $100K+. Anduril charges millions.
Any platform
UGVs, FPV drones, fixed-wing UAVs, multirotor platforms, or fixed installations. One AI system, every form factor. MAVLink, PX4, and DJI compatible.
Zero cloud dependency
Everything runs on-device. Detection, tracking, retraining, targeting. Works in DDIL environments where connectivity is denied, degraded, or non-existent.
Real-time inference
YOLOv8 + BoT-SORT multi-object tracking at 15+ FPS with tracking, 30+ FPS with TensorRT. Low enough latency for FPV terminal guidance.
Human in the loop
Operators classify, correct, and designate targets through an intuitive web UI. RC override at any time. E-STOP always available. AI assists, humans decide.
One system. Every platform.
RAPTOR deploys across the full spectrum of unmanned and fixed systems.
Ground
Unmanned Ground Vehicles
Autonomous threat detection and target following for ground robots. MAVLink integration steers toward designated targets. RC override at any time.
- ✓ Auto-steer toward designated targets
- ✓ ArduRover / ArduPilot compatible
- ✓ Manual/Guided mode switching via RC
Air · Rotary
FPV Drones
AI-powered terminal guidance for first-person-view platforms. Lock a target, and RAPTOR maintains tracking through jamming and signal loss—the capability Ukraine proved in combat.
- ✓ Visual target lock survives RF jamming
- ✓ 60g Jetson module fits 5" frames
- ✓ Lightweight headless mode for payload constraints
Air · Fixed-wing
Fixed-Wing UAVs
Persistent ISR with automatic detection alerts. RAPTOR watches the feed so the operator doesn't have to stare at a screen for hours. Aerial-perspective models trained on VisDrone and DOTA datasets.
- ✓ Auto-alert on detection events
- ✓ Bandwidth-adaptive streaming
- ✓ Loiter pattern generation
Fixed site
Fixed Installations
Perimeter security, checkpoint overwatch, and FOB protection. Multi-camera coverage on a single Jetson. 24/7 operation with thermal/IR for night.
- ✓ Multi-camera array management
- ✓ Day/night auto-switching models
- ✓ Alert escalation and alarm integration
One loop: detect, classify, act, retrain
Every stage feeds the next. What the operator corrects today is what the model knows tomorrow.
1 · Detection & Tracking
- ✓ Real-time YOLOv8 inference with TensorRT acceleration
- ✓ BoT-SORT multi-object tracking with persistent IDs (300-frame memory)
- ✓ Dual-model detection: primary + COCO classes simultaneously
- ✓ Auto-detect armed persons via weapon-person proximity
- ✓ Adjustable confidence threshold (live slider)
2 · Classification & Labeling
- ✓ Tag detections: THREAT / FRIENDLY / NEUTRAL / UNKNOWN
- ✓ Correct misclassifications in real-time from the UI
- ✓ Labels persist across restarts
- ✓ Identity reassignment when track IDs change
- ✓ Audio alerts on new threat detection
3 · Targeting & Engagement
- ✓ Crosshair overlay on designated targets
- ✓ Multi-target priority queue with auto-designation
- ✓ Autonomous target following (UGV steering)
- ✓ E-STOP: spacebar or button, instant disengage
- ✓ RC override always available—operator stays in control
4 · Field Retraining
- ✓ Auto-save labeled data on every correction
- ✓ One-click retrain from the web UI
- ✓ Custom class list grows as operators add corrections
- ✓ Import external datasets (VisDrone, DOTA, xView)
- ✓ Batch annotation review and quality control
Detections on every screen
RAPTOR is being built to push Cursor on Target (CoT) events directly into the TAK ecosystem, so every threat appears on every operator's ATAK display in real time—no manual reporting, no radio calls, no delay.
Auto-Populate SA
Detected threats, vehicles, and persons of interest appear as CoT markers on ATAK maps. Classified as hostile, friendly, neutral, or unknown with confidence scores.
Works Disconnected
RAPTOR pushes CoT over local UDP—no internet required. Works with any TAK Server on the local mesh, including OpenTAKServer running on a companion device.
Multi-Node Coverage
Deploy three RAPTOR nodes around a compound. Each pushes detections to TAK independently. Operators see a unified threat picture from all sensors on a single map.
How RAPTOR compares
| System | Cost | Edge Native | Field Retrainable | Platform Agnostic |
|---|---|---|---|---|
| RAPTOR | $250 compute | ✓ | ✓ | ✓ |
| Anduril Lattice | $millions | ✓ | ✗ | ✗ |
| Shield AI | $100K+/unit | ✓ | ✗ | ✗ |
| Skydio | $10K+/drone | ✓ | ✗ | ✗ |
| Palantir AIP | $5M+/yr | ✗ | ✗ | N/A |
Technical details
Compute
- NVIDIA Jetson Orin Nano / Orin NX
- 8GB–16GB unified memory
- JetPack 6 with CUDA + TensorRT
- 60g module weight
Performance
- 15+ FPS (detection + tracking)
- 30+ FPS with TensorRT
- <30ms inference latency
- 300-frame track persistence
Interface
- Web UI (any device on network)
- MJPEG annotated video stream
- MAVLink vehicle control
- systemd service (boot-ready)
Where RAPTOR is today
The capabilities above run in the current build. The items below do not—we would rather list them as roadmap than imply them as product. Status labels are exact.
ATAK / CoT Integration
Push detections as Cursor on Target messages into the TAK ecosystem via UDP. Threat detections auto-appear on every ATAK device in the network—full battlefield integration with zero manual reporting.
Multi-Source Video Input
RTSP IP cameras, video files, and live webcam feeds. Drag-and-drop video upload from the web UI. Switch sources without restarting. Test and validate with recorded footage.
Automated Evidence Recording
Auto-record video clips when threats are detected. Timestamped incident recordings stored locally, downloadable from the web UI. Full chain-of-custody for after-action review.
TensorRT Optimization
Hardware-accelerated inference using NVIDIA TensorRT. 3–5× FPS improvement over PyTorch, enabling 30+ FPS real-time detection with full tracking on the $250 Jetson module.
Ceradon Sim Integration
Virtual testbench with ArduPilot SITL and Gazebo 3D. Test RAPTOR + UGV control in simulated environments before deploying to real hardware. Sim-to-real parameter export.
Gimbal Control
Slew-to-cue: automatically point camera at detected threats. SIYI, Gremsy, and MAVLink gimbal protocol support.
NanoTrack Terminal Guidance
Single-object tracker optimized for FPV terminal guidance. 60 FPS on Jetson. Operator locks target, AI maintains track through occlusion, scale change, and motion blur.
Thermal / IR Camera
FLIR Lepton, Seek Thermal, and FLIR Boson integration for 24/7 night operations.
Multi-Node Mesh
Coordinate multiple RAPTOR nodes across a perimeter. Shared detection data, deconfliction, and distributed tracking over BATMAN mesh networking.
Multi-Sensor Fusion
Combine RAPTOR (visual), Vantage (passive RF through-wall), and RF detection (RTL-SDR) into a single $300 multi-phenomenology sensor node. Nobody else offers this at any price.
GPS-Tagged Detections
Lat/lon/alt for every detection via MAVLink telemetry. Geofenced alert zones and geospatial overlay.
Ready to deploy intelligent vision?
Whether you're integrating AI onto a UGV, adding autonomy to FPV platforms, or building a perimeter security system—RAPTOR gives you field-retrainable AI at a price point that makes sense.
SDVOSB · CAGE 179U9 · UEI UZA9PFJ9RDL6