TinyDEVO: Deep Event-based Visual Odometry on Ultra-low-power Multi-core Microcontrollers

A team from ETH Zurich has developed TinyDEVO, a breakthrough deep learning model that brings sophisticated visual odometry (VO)—the AI that estimates camera motion—to ultra-low-power microcontrollers. By combining neural network architectural optimizations and hyperparameter tuning, they dramatically compressed the leading event-based VO model, DEVO. The result is an 11.5x reduction in memory footprint (down to 63.8 MB) and a 29.7x cut in computational operations (to 5.2 billion MACs per frame), with only a modest increase in trajectory error.

This efficiency breakthrough allowed the team to deploy TinyDEVO on a commercially available, ultra-low-power 9-core RISC-V microcontroller. The system runs at approximately 1.2 frames per second while consuming a mere 86 milliwatts of power. This demonstrates, for the first time, the feasibility of running a complete event-based VO pipeline directly on a resource-constrained MCU, a feat previously impossible due to the massive memory and compute demands of state-of-the-art models.

The work is significant because it directly addresses a critical bottleneck in autonomous edge devices. Visual odometry is a core component for navigation in robots, drones, and augmented reality glasses, but traditional implementations are too power-hungry for long battery life. TinyDEVO's microcontroller-level efficiency opens the door to a new generation of tiny, intelligent, and energy-autonomous machines that can understand and navigate their environment without relying on cloud connectivity or large batteries.