AUTONOMY · OBSERVABILITY · RESILIENCE

A mission pipeline.
Evidence at every boundary.

Virtual sensors, CPU ONNX inference, localization and planning, supervised by C++17. Explore what happens when data or the companion process disappears.

Loading evidence
01 / Boot gate02 / Camera + LiDAR + IMU + GNSS03 / ONNX CPU04 / Localization + A*05 / C++ supervisor06 / Autopilot stub
SCENARIO RESULT——
ONNX INFERENCE · P95—WSL CPU wall-clock measurement
SUPERVISOR IPC · P95—Python ↔ C++ round trip
LOCALIZATION RMSE—Estimated vs. simulated position

Mission replay

INIT
Ground truthEstimateA* routeObstacles
0.0 s

Perception & health

Actual synthetic RGB frame consumed by ONNX

64 × 48 synthetic RGB · hand-authored brightness graph · no trained threat classifier

Supervisor reason—
GNSS age—
Position error—
Position covariance trace—
Battery (simulated)—

State transitions

Boot & artifact trust

Signature checks run in user space with an ephemeral demo key. Hardware trust anchors and fuse lifecycle are modeled only.

Update verification experiments

This is an accelerated kinematic simulation. The replay shows recorded runs, not a live vehicle. This particular replay uses the fast policy harness. The separate PX4 flight replay demonstrates actual ROS 2 and Gazebo/PX4 integration. Qualcomm boot ROM, NPU execution and RF remain outside scope. The ONNX model detects a bright synthetic object; it cannot identify a hostile drone.
Reproduce the results →Inspect the design →Domain notes →Measured evidence →