Research demo — not deployment-ready.
SIH26168 · AI-ML based Intelligent Dead Reckoning · ISRO · Smart Vehicles
Your phone loses GNSS because a roof got in the way, not because the satellites failed. StillDot keeps a position estimate using only the IMU the phone already has. No OBD wire. No internet. No actuation.
The network estimates speed and uncertainty. The filter estimates position. We never integrate acceleration twice to get distance.
The pitch script says: if the phones fail, replay on the laptop. That is this console.
uv sync --group dev
cd frontend && npm install && npm run build && cd ..
uv run stilldot demoOpen http://127.0.0.1:8000. Tap START. The room-walk path draws
live, then the drift figure is read out in metres and as a percentage
of distance travelled. The requirement is under 10%.
Airplane mode on a phone is the live path. This laptop replay is the same engine on a surveyed synthetic path — labelled as such on screen.
| Built and measured | Not built |
|---|---|
| Time-aligned IMU, alignment, virtual odometer, ZUPT, NHC, 10 Hz output | IO-VNBD training / country hold-out numbers |
| Room walk (50 m) and tunnel (1 km / 60 km/h) replays | Full invariant EKF on a Lie group |
| Naive ∫∫a comparison (the t² lesson) | Map matching against OpenStreetMap |
| Live DeviceMotion ingest | A published Indian-road score |
Do not quote an accuracy number without the scenario, the distance, and the fact that these replays are synthetic.
uv run ruff check src tests
uv run mypy src/stilldot
uv run pytest
cd frontend && npm run checkNever say “we used AI to predict the position.” Say: the network estimates speed and uncertainty; the filter estimates position.
Apache License 2.0. See LICENSE.