[AIoT] [Neurusie] SIGMA

Inspiration

Intersections are dumb. Rigid light cycles waste time on empty lanes, and when a crash happens, the system has no idea. The first minutes after an accident decide lives. We wanted an intersection that actually understands what's happening on it.

What it does

SIGMA runs in two modes. Daily: adapts lights to traffic flow — longer green where cars wait, faster red for empty lanes. Emergency: detects crashes by sound (crash, tire squeal, glass, sirens), switches all lanes red, triggers LED alarm, sends Wi-Fi alert.

How we built it

Two Nordic boards. nRF54L15-DK with a Time-of-Flight laser sensor measures traffic. nRF7002-DK is the brain — TinyML and LED matrix signalization. Everything hand-soldered. Two AI models. An RL policy trained on UXSim simulator with reward-shaping, exported as a C-header and flashed straight to the nRF7002 — on-device inference, no cloud. A CRNN from AudioSet (cnn8rnn-audioset-sed) for accident sound detection.

Challenges we ran into

Stable real-time comms between the two boards took longer than expected. Hand-soldering needed some overnight fixes. Shrinking the RL policy to fit nRF7002 memory while keeping decision quality took several quantization iterations. Balancing audio detection speed with false-positive prevention was tricky.

Accomplishments that we're proud of

A working end-to-end system: sensor → AI → signalization. RL policy trained to sensible behavior and actually flashed to embedded target. All our hardware works. And SIGMA is a real acronym.

What we learned

Integration is harder than individual components. TinyML on a tiny target means designing for memory and latency from day one, not at export. Well-shaped rewards separate a policy that works from one that looks like it does.

What's next for [AIoT][Neurusie] SIGMA — Smart Intersection Guardian of Midgard's Avenues

A mesh of intersections talking to each other — a warband instead of a lone guardian. Green wave for ambulances along their full route. Our RL code already supports any number of arms, so scalability is built in. Long-term: integration with city traffic management as a Smart City edge layer.

Built With

• ble
• c
• coffee
• nrf45l15-dk
• nrf7002-dk
• threads