TinyDEVO on microcontrollers

Visual odometry is the job of figuring out how a camera moved by comparing what it sees over time. On April 9, 2026, researchers posted TinyDEVO, a version built to run on an ultra-low-power microcontroller instead of a graphics processor. (arxiv.org) The paper is by Alessandro Marchei, Lorenzo Lamberti, Daniele Palossi, and Luca Benini of ETH Zürich, IDSIA at Università della Svizzera italiana and SUPSI, and the University of Bologna. They describe TinyDEVO as a deep event-based visual odometry model for resource-constrained microcontroller units. (arxiv.org) An event camera does not send full images like a phone camera. Its pixels report only brightness changes, which cuts redundant data and helps in fast motion, motion blur, and difficult lighting. (arxiv.org) That matters because earlier deep event-based visual odometry systems were too large for tiny edge chips. The TinyDEVO paper says the earlier DEVO model needed 733 megabytes and 155 billion multiply-accumulate operations per frame, while TinyDEVO cuts that to 63.8 megabytes and 5.2 billion operations per frame. (arxiv.org) The team ran TinyDEVO on a nine-core RISC-V microcontroller and reported about 1.2 frames per second at an average power draw of 86 milliwatts. In the same paper, they identify the chip as a GAP9 system-on-chip, a part aimed at low-power edge machine learning. (arxiv.org, github.com) The tradeoff is accuracy. The authors report an average trajectory error of 27 centimeters, which they say is 19 centimeters worse than DEVO across three state-of-the-art datasets. (arxiv.org) TinyDEVO follows a line of work that tried to make event-only motion tracking practical without adding inertial sensors, stereo cameras, or regular frame cameras. The original DEVO paper, submitted in December 2023 and accepted to the International Conference on 3D Vision 2024, said those extra sensors add cost and system complexity. (arxiv.org, github.com) The new paper places that idea inside a much tighter power budget. Its table compares TinyDEVO at roughly 0.09 watts on GAP9 with DEVO measured at 27.5 frames per second on a GeForce RTX 4070 and other visual odometry systems running on larger processors such as Jetson Xavier AGX and Intel Core i7 hardware. (arxiv.org) The authors frame the target devices as battery-limited embedded systems, including wearables, miniature robots, and fixed sensors. Their claim is narrower than replacing high-end robotics computers: they say the result shows event-based visual odometry is feasible on ultra-low-power devices for the first time. (arxiv.org) What comes next is the usual edge-computing question: whether 1.2 frames per second and a 27-centimeter average error are enough for a real product. TinyDEVO does not settle that on its own, but it moves camera-motion estimation onto hardware that runs in the tens of milliwatts, not the hundreds of watts. (arxiv.org)