TinyDEVO: event‑based VO on MCUs
TinyDEVO demonstrated deep event‑based visual odometry that runs on ultra‑low‑power microcontrollers, pointing to perception approaches suitable for constrained robotic platforms. The work targets energy‑efficient on‑board perception for small or battery‑limited robots. (x.com)
Visual odometry is the job of estimating how a camera moves through space from what it sees. A paper posted April 9 says that kind of motion tracking can now run on an ultra-low-power microcontroller instead of a graphics processor. (arxiv.org) The system is called TinyDEVO, for Tiny Deep Event-based Visual Odometry, and it was written by Alessandro Marchei, Lorenzo Lamberti, Daniele Palossi, and Luca Benini of ETH Zürich, IDSIA, USI-SUPSI, and the University of Bologna. The authors say they deployed it on a nine-core RISC-V microcontroller and measured about 1.2 frames per second at an average power draw of 86 milliwatts. (arxiv.org) The camera in this setup does not record ordinary video frames. It uses an event camera, which reports only brightness changes at each pixel, a design that can hold up better under motion blur and difficult lighting than standard frame cameras. (arxiv.org) That matters for small robots, wearables, and battery-limited devices that cannot carry a 30-watt graphics module. In the TinyDEVO paper’s comparison table, the earlier Deep Event Visual Odometry system ran at 27.5 frames per second on an Nvidia RTX 4070 and used 733 megabytes of memory, while TinyDEVO is listed at 108 frames per second on the same class of graphics card and 1.2 frames per second on the microcontroller. (arxiv.org) The main tradeoff is compression. The authors report cutting memory use 11.5 times, to 63.8 megabytes, and reducing compute 29.7 times, to 5.2 billion multiply-accumulate operations per frame, compared with Deep Event Visual Odometry. (arxiv.org) Accuracy fell, but not by as much as the hardware budget did. On three state-of-the-art datasets, the paper reports an average trajectory error of 27 centimeters, which it says is 19 centimeters worse than Deep Event Visual Odometry. (arxiv.org) Deep Event Visual Odometry itself was introduced in December 2023 and accepted to the 2024 International Conference on 3D Vision. Its authors said it was the first monocular, event-only system with strong results across multiple real-world benchmarks and that it cut pose-tracking error by as much as 97 percent against earlier event-only methods. (arxiv.org) TinyDEVO shifts that line of work toward hardware that fits inside embedded systems. The paper says its microcontroller target is a GAP9 system-on-chip, a RISC-V design aimed at low-power machine learning and signal processing at the edge. (arxiv.org) The authors describe the result as the first demonstration of an event-based visual odometry pipeline on an ultra-low-power device. If that claim holds up under follow-on testing, it puts full on-board motion estimation closer to the power envelope of tiny flying robots and always-on sensors. (arxiv.org)
