Kestrel

Watch a real mission

This viewer plays back a real recorded flight. It is not an animation. The flight path and mission states come from scripts/record_replay.py, run alongside a live simulated mission.

Both recordings use the shipped YOLO detector at its production confidence threshold. Live in-sim detection has a known limitation, a domain gap between real training photos and simulated textures. Neither recording found a confirmed defect. This is the honest, measured behavior of the shipped system, not a demo shortcut. See a sample report with confirmed defects from an earlier ArUco run and the benchmarks table below for real detection numbers.

How it works

These steps happen during one mission.

1. The mission planner builds a survey path around the structure.
2. The flight commander flies the copter to each waypoint.
3. The defect detector reads the camera feed during the whole flight.
4. On a detection, the mission director pauses the survey and flies a closer orbit around the defect.
5. The system stores each photo and each detection.
6. After landing, the report writer writes a report in markdown format.

Benchmarks

These numbers come from scripts/score_mission.py, run against real mission data. Detection rate is the count of defects found over the count of ground truth markers. Localization error is the distance from each detected defect to its nearest ground truth marker.

The yolo row counts raw detections against the ground truth count. This number does not confirm that each detection matches a real marker. Only one of the four detections in that row was a true marker, 0.37 m from ground truth. See the full benchmarks file for the complete explanation.

Detector model

The corrosion detector is a YOLOv8n model, trained on the public corrosion-bi3q3 data set (CC BY 4.0) and deployed with ONNX Runtime. The script scripts/eval_model.py scores the shipped model against the held out test split through the same inference path that the flight stack uses. The test split played no part in training or in model selection.

The test score matches the validation score (0.617), so the model did not overfit to the validation split, and the ONNX export gives the same quality as the training checkpoint. See TRAINING.md for the training and evaluation steps.

Onboard intelligence

The copter runs several AI systems onboard. Every model exports to ONNX and runs with ONNX Runtime on the CPU, the same target a small flight computer uses. Every system is advisory. The safety guard node keeps final control.

The health monitor is a small autoencoder trained on normal flight telemetry. It fires on a sustained anomaly, not on a single noisy reading, and it fires on an injected motor fault. The model learns normal from data, so it must be trained on telemetry from the same conditions it runs in. See the health benchmarks.

The models are measured for the edge. INT8 quantization shrinks the detector from 11.7 MB to 3.2 MB, and the mAP50 changes only from 0.621 to 0.613. See the edge benchmarks.

Talk to it

The flight stack also works as an MCP server. An LLM client can fly the simulation through this server. The client can ask for telemetry data. The client can send a takeoff command. The client can send the copter to a position. The client can send a land command. Every command goes through the same guarded services that the autonomous mission uses. The safety guard node keeps final control at all times.

See the README for setup steps and the full tool list.

Quick start

git clone https://github.com/ajnieves1/Kestrel.git
docker compose -f docker/compose.yaml build dev
docker compose -f docker/compose.yaml run --rm dev ros2 launch kestrel mission.launch.py headless:=true