TinyOdom: Hardware-Aware Efficient Neural Inertial Navigation.

Area of Science:

• Robotics and Autonomous Systems
• Machine Learning for Navigation
• Embedded Systems Engineering

Background:

• Deep inertial odometry offers high-resolution trajectory estimation in GPS-denied areas.
• Existing neural inertial dead-reckoning systems are too resource-intensive for ultra-resource-constrained (URC) devices.
• Current methods face challenges like gravity pollution, sensor disturbances, and altitude estimation failures.

Purpose of the Study:

• To develop TinyOdom, a framework for training and deploying lightweight neural inertial models on URC hardware.
• To enhance robustness against environmental and sensor perturbations for accurate dead-reckoning.
• To enable real-time, high-performance inertial odometry on devices with limited memory, power, and computation.

Main Methods:

• Utilized hardware-aware and quantization-aware Bayesian neural architecture search (NAS) with a temporal convolutional network (TCN) backbone.
• Introduced a novel magnetometer, physics, and velocity-centric sequence learning formulation.
• Expanded 2D to 3D learning with a model-free barometric g-h filter for robust altitude estimation.

Main Results:

• TinyOdom achieved model size reductions of 31× to 134× across diverse applications (pedestrian, animal, aerial, underwater).
• Demonstrated localization accuracy with 2.5m to 12m error in 60 seconds, outperforming state-of-the-art methods.
• The barometric filter maintained altitude tracking within ±0.1m, robust to disturbances.

Conclusions:

• TinyOdom facilitates direct deployment of advanced neural inertial odometry on URC devices.
• The proposed sequence learning and barometric filtering techniques significantly improve localization performance and robustness.
• This work bridges the gap between high-performance inertial odometry and the constraints of embedded systems.