New AI Tracks Where You Walk Even When GPS Dies

GPS is a modern miracle until it isn't. Step inside a mall, a tunnel, a parking garage or a dense city street, and the satellite signal fades out — leaving your phone, a delivery drone or a rescue robot with no idea where it is. So researchers have turned to the tiny motion chip already in almost every device: an accelerometer and gyroscope, together called an IMU, that feel every tilt, turn and bump. The idea is to let AI read those wiggles and calculate how far you've walked. The problem is that these systems have been frustratingly fragile.

The weakness is orientation. A sensor is a small chip, and it can be mounted at any angle — sideways in a phone, upside down in a drone, rotated in a fitness tracker. Older AI systems quietly assumed a fixed angle, so a simple twist could wreck their accuracy. GINIO fixes this by building the math of rotation directly into the model, so its answer rotates correctly along with the sensor. On a standard walking test, GINIO's average position error was about 2 meters; a comparable method blew up to 76 meters. It also needed roughly 12 times less computing power.

On a drone test, a GINIO-style version cut error from 5.6 meters to 1.4 meters using only onboard sensors. Most striking: on a dataset where a sensor was physically unplugged and re-mounted at a new angle, the older approach's error jumped to 8.15 meters, while GINIO stayed near 0.5 meters — without any retraining. The team also gave the AI a way to say "I'm not sure," reporting its own uncertainty, which makes it far safer to plug into real navigation software.

The catch: this is a research paper, not a product on sale. It was tested on recorded datasets and drone flights, not on millions of phones in the wild, and motion sensors still drift over time — meaning GINIO is best used alongside other clues like cameras or WiFi, not as a GPS replacement on its own. Expect the first real-world wins in robots, drones and augmented-reality headsets before your phone.