Tracking Large-scale Shared Bikes with Inertial Motion Learning in GNSS Blocked Environments

Researchers propose an inertial motion learning framework for tracking shared bikes in GNSS-denied environments like urban canyons, combining mechanical constraints with mixture-of-experts models to achieve 12% accuracy improvements over baselines. The system leverages pedaling behavior patterns to dynamically calibrate wheel speed estimates, demonstrating practical viability through real-world deployment data from DiDi's bike-sharing platform.

This research addresses a fundamental challenge in urban mobility infrastructure: reliable positioning of shared bikes in environments where satellite signals fail. Urban canyons, tunnels, and dense building structures render traditional GNSS useless, yet tracking accuracy directly impacts operational efficiency, user experience, and asset management for bike-sharing platforms. The proposed solution cleverly sidesteps expensive sensor deployments (LiDAR, cameras) that would be economically unfeasible at scale, instead extracting actionable positional data from low-cost inertial measurement units already embedded in most modern bikes.

The technical innovation centers on two insights: mixture-of-experts architectures enable adaptive weighting of different motion models, while mechanical constraints derived from bicycle physics create deterministic bounds on motion. By analyzing the cyclic relationship between pedaling frequency and wheel acceleration, the framework performs self-calibration without external references. This represents meaningful progress in sensor fusion for edge computing, as it reduces infrastructure dependencies while improving robustness.

For the shared mobility industry, the implications are significant. Reduced localization errors directly translate to improved dispatch algorithms, theft prevention, and maintenance routing. The 12% accuracy improvement from baseline methods could meaningfully reduce operational costs and enhance user experience through better real-time positioning. DiDi's involvement validates commercial relevance—large platforms depend on reliable tracking for operations across multiple cities.

Future development should focus on generalization across different bike models and riding styles, temperature compensation for inertial sensors, and integration with cellular-based positioning for hybrid coverage scenarios. Adoption could accelerate the transition toward truly autonomous mobility systems that operate reliably in complex urban environments.

• →Mixture-of-experts model with mechanical constraints achieves 12% accuracy improvement for bike tracking without external infrastructure in GNSS-denied areas
• →System extracts wheel speed calibration from pedaling behavior patterns, enabling self-correction without manual recalibration
• →Low-cost inertial sensors provide viable alternative to expensive LiDAR/camera systems for large-scale shared bike deployment
• →Real-world validation from DiDi platform demonstrates commercial viability for urban mobility applications
• →Technology reduces infrastructure dependencies while maintaining robust positioning in complex environments