SmartPilot: A Lightweight, Low-Power Edge AI Module for Intelligent UAV Operations

Small UAVs are increasingly relevant to security and defense missions, yet onboard autonomy remains constrained by mass, power, latency, and integration complexity. We present a 22 g edge-AI module, with an accelerator, perception network, and control stack co-designed for a 6 W peak budget, that adds human-supervised perception, tracking, and target-directed control to fixed- and rotary-wing UAVs without replacing the existing flight controller.

A purpose-designed joint detection-and-embedding (JDE) network detects and associates targets in a single inference pass. A reinforcement-learning policy developed and trained in simulation converts its outputs into thrust and roll-, pitch-, and yaw-rate commands. The module therefore occupies the pilot's role rather than the flight controller's, issuing the same high-level commands a human pilot would. The operator retains supervisory authority and immediate override.

The INT8-quantised perception network occupies 0.73 MB and executes in 12 ms at 320 × 320, with 15.2 ms of p99 latency; quantisation reduces mAP@0.5 by only 0.009. Detection achieves 0.746 mAP@0.5 in-domain and 0.614 on the AOT benchmark. Input mapping dominates quantisation error; whole-frame letterboxing reduces AOT mAP@0.5 to 0.127, whereas native-scale cropping restores 0.614. 

Closed-loop evaluation covered 350 engagements in a DCS-based simulator, spanning 14 target types and 25 approach geometries, with the physical module commanding an unmodified commercial flight controller. Averaged over flights, field-of-view retention over measurable frames and correct-identity tracking over visible frames both reached 94.7%. The closing rate exceeded 1 m/s for 77.2% of frames, and 47.7% of engagements ended in a successful intercept. Flight trials on physical airframes await authorisation from national authorities. 

Hardware–network co-design brings a complete perception-to-control layer within the mass and power envelope of small UAVs, without displacing the existing flight controller.