AI Business

A closed-loop learning architecture linking the simulation environment and the RL model through state, reward, and action.

• PyGame-based repeatable test environment providing state and reward in real time
• Seven observations: flight stage · yaw · pitch · tilt · target screen coordinates (dx, dy) · elapsed time (dT)
• Three actions: yaw · pitch · tilt (23 discrete levels each, mapped to angular rates)
• Reward design: per-frame proximity to image center plus terminal hit/miss rewards
• Coordinate system defined by azimuth, elevation, line of sight (LOS), and field of view (FOV)

A four-stage flight scenario — boost, cruise, synchronize, attack — with terminal decision logic.

• Randomized initial conditions: speed 100–200 m/s · initial pitch −40 to −10° · boost 3–5 s
• Randomized target speed of 30–60 km/h
• Automatic switching of attitude and camera control targets (pitch · tilt) per stage
• Terminal outcomes: HIT · CRASH · loss of lock · MISS
• Training monitor: camera view · map · altitude and attitude graphs

Actor/Critic policy learning with PPO (Proximal Policy Optimization) on Stable-Baselines3.

• Implemented on Stable-Baselines3, the de-facto standard RL library
• Actor/Critic architecture with two hidden (linear) layers
• Actor [128, 128] · Critic [256, 256] networks
• Convergence analysis comparing learning rates of 0.001 and 0.0003

Real-time object detection and target lock-on maintenance from monocular RGB video.

• Bounding-box output from deep-learning detectors (YOLO and MobileNet families)
• Target selected as the detection whose center is nearest the lock-on position
• When objects cross or overlap, the detection closest to the predicted position is kept as the same target
• Loss of lock is declared when the target leaves the frame beyond a set duration

Converting image-coordinate error into line-of-sight (LOS) angular error to generate guidance and camera commands.

• Error between target center and image center, normalized by frame size
• Conversion to horizontal and vertical LOS error using HFOV and VFOV
• Real-time generation of vehicle yaw/pitch and camera tilt/zoom commands
• Communication protocol design across simulator, controller, and detector

Real-time inference verified on a Raspberry Pi 5 with a domestic DeepX NPU (dx-m1).

• YOLOX-Nano averages 6.5 ms — roughly 10× faster than the same board's CPU (69.2 ms)
• On the NPU: YOLOX-Tiny 8.2 ms · MobileNetV2+SSD 8.4 ms · YOLO26n 19.6 ms
• Benchmarked against PC GPU/CPU environments, confirming embedded real-time feasibility
• Verified against a Unity-based flight simulator — target range 1.5–2.5 km · speed approx. 170 m/s

The hostile loitering munition is first detected 2 km out and the engagement is authorised at the control station.

• Low-altitude radar first detects the intruding loitering munition at 2 km
• Detected threat-track data is transmitted to the control-station system
• The operator confirms the threat on the display and authorises the intercept
• On authorisation the interceptor drone launches toward the predicted target point

Digitron's own AI algorithms analyse the interceptor drone's EO camera video in real time.

• The drone begins its manoeuvre and activates its onboard electro-optical camera
• In-house AI algorithms analyse the live EO video stream
• The algorithm identifies the visual signature of the hostile drone
• Real-time detection draws a precise target box, then the system switches to tracking mode

The drone computes its own intercept trajectory from the AI detection data and the target's flight path.

• Optimal intercept trajectory computed onboard from AI detection data and target motion
• Precision guided flight along the computed trajectory to close on the hostile aircraft
• Neutralisation of the loitering munition in the air