AI-ML based Intelligent Dead Reckoning system for seamless navigation
Indian Space Research Organisation(ISRO) · Miscellaneous · Software
A named benchmark, a real approach and a relatable, distinctive problem — build the learned drift correction and clean re-blending, and be honest that accuracy degrades with blackout length, because the physics is unforgiving and an ISRO judge knows it.
Data: IO-VNBD inertial and odometry benchmark (github.com/onyekpeu/IO-VNBD)
What it actually is
Navigation apps freeze or jump when GPS drops — in tunnels, multi-level car parks, urban canyons — and most Indian vehicles have only the driver's smartphone, not a factory inertial system wired to the wheels. The ask is an AI dead-reckoning system that uses the phone's own accelerometer and gyroscope to keep tracking position through a GPS blackout, despite the noise of a phone on a dashboard.
What to build
A dead-reckoning module that, during GPS outage, estimates vehicle position from a smartphone's MEMS accelerometer and gyroscope alone — no OBD-II speed feed — learning to correct the sensor biases, drift and vibration noise that make naive integration diverge, using the IO-VNBD benchmark to train and validate, detecting outage entry and re-acquisition to blend cleanly back to GPS-aided tracking, and holding position error to a usable level over the tens of seconds to few minutes a real tunnel or urban-canyon blackout lasts.
Smallest thing that wins the room
Replay an IO-VNBD sequence with GPS artificially cut through a tunnel-length section and show your AI dead-reckoning holding the estimated track close to ground truth while naive double-integration diverges rapidly, with the position error at re-acquisition reported.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 36% of the 226 · #145 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: defence, intelligence and space bodies drew small fields.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
4/5A strong pick — the named public benchmark means you can train and validate rigorously, the problem is genuinely useful and relatable, learned inertial odometry gives you a real approach, and the phone-only constraint is a distinctive and honest framing, though you must be candid that accuracy degrades with blackout length.
Feasibility
3/5The IO-VNBD benchmark is public and named, and learned inertial odometry is an established research direction, but dead reckoning from a phone IMU alone without any speed reference is fundamentally hard — errors accumulate quadratically — so holding useful accuracy beyond a short blackout is a real, unsolved-in-general challenge.
Innovation scope
4/5Learning to suppress phone-specific vibration and bias to extend usable dead-reckoning time from a phone IMU alone is genuinely open, and the outage-detection and re-blending logic is your own design.
Clarity
4/5The description defines the problem, the sensors, the no-OBD constraint and the target environments precisely, and names the benchmark dataset, so the target is clear even though no accuracy threshold is stated.
Effort
HeavyThe learned inertial model, the outage detection and re-acquisition blending, and validation on the benchmark are focused, well-bounded work on data that already exists.
Demo-ability
MediumYour track holding while naive integration diverges is a clear, satisfying result, but it plays out on a plotted trajectory that needs the GPS-outage context to be appreciated.
In its favour
- Green flag: IO-VNBD is a named public benchmark built for exactly this ground-vehicle positioning problem, so you can train and validate on real data
- Green flag: Learned inertial odometry approaches like RoNIN give you an established, effective starting point rather than a from-scratch model
- Green flag: The phone-only, no-OBD constraint is a distinctive framing that matches the reality of most Indian vehicles
- Green flag: The GPS-drops-in-a-tunnel problem is instantly relatable to any judge, so the motivation needs no explanation
Against it
- Red flag: Dead reckoning error grows quadratically with time, so accuracy over a long blackout is fundamentally limited and overclaiming it is exposable
- Red flag: A phone IMU without a speed reference cannot resolve velocity scale well, which is exactly why the no-OBD constraint makes this hard
- Red flag: Vibration and orientation changes from a phone loose in a holder inject noise that a model trained on cleaner data will not handle
- Red flag: An ISRO navigation judge knows the physical limits precisely and will ask how accuracy scales with outage duration
What you will be writing
- Learned inertial odometry (IONet / RoNIN-style)
- IO-VNBD benchmark dataset
- MEMS IMU bias and drift modelling
- GPS outage detection + re-acquisition blending
- Sequence models (LSTM / TCN) on IMU streams
- Trajectory error evaluation (ATE / RTE)
- Inertial navigation
- Dead reckoning
- Sensor fusion
Prior art to read before you start
smartphone inertial odometry · GPS-denied dead reckoning · learned IMU drift correction
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.