AEGIS

A Computer-Vision Turret,
From First Principles

A pan/tilt turret that sees, tracks and aims through the full perception → control → fire-control loop. The interesting parts (PID control, Kalman and α-β filtering, ballistic intercept, stereo depth, optimal assignment and a convolutional net) are hand-written from scratch and tested against references. Every demo below runs that real Python in your browser via Pyodide.

Aim

Control & Safety

PID Control & the Safety Gate

Tune the real PID loop driving the gimbal onto a moving target, then exercise the live safety gate that fires only on its allowlist.

Open demo →

Range & Fire

Fire-Control

Stereo Fire-Control

Recover range from a stereo pair, then solve a ballistic intercept (lead, gravity hold-over and drag) so dart and target meet.

Open demo →

See

Vision · CNN

A CNN, Written From Scratch

A convolutional net built in pure NumPy (conv, pool, FC and the backprop that trained it, zero autograd), classifying patches live.

Open demo →

Estimate

Algorithms

Algorithms From Scratch

Kalman filtering, the Hungarian assignment algorithm, non-max suppression and block-matching stereo, each from first principles.

Open demo →

Built for the build. AEGIS targets an NVIDIA Jetson Orin Nano. The whole pipeline is validated in simulation with a fast pure-Python test suite; the bug-prone maths is written without heavy dependencies so it can be unit-tested, and bundled unchanged into the demos above by tools/build_site.py.

Further work

Physical build: pan/tilt gimbal, flywheel gun and stereo cameras on the Jetson Orin Nano; bring-up of the same control loop against real servos.
Real dataset: capture and label real footage, train the custom YOLO detector, and deploy it as a TensorRT engine on the Jetson.
Closing the loop: the full perception → tracking → fire-control chain running end-to-end on hardware, with the sim-tuned gains as the starting point.