Nessie Circuits Shows Off Riotee's Battery-Free Chops with a Solar-Powered Hot-Word Neural Network

Nessie Circuits' Kai Geissdoerfer and Marco Zimmerling have collaborated on a paper showcasing the company's Riotee, an open source platform for battery-free Internet of Things (IoT) devices — showing off use-cases including an on-device tiny machine learning (tinyML) deep neural network (DNN) for batteryless solar-powered hot-word detection.

"The rapidly growing Internet of Things ( IoT) can avoid the high cost and environmental burden of replacing trillions of batteries by using sustainable battery-free devices that operate maintenance-free for decades," Geissdoerfer and Zimmerling write in their paper on the topic. "To develop battery-free IoT systems, researchers and makers require a common platform that is versatile, affordable, and easy to use. However, limited availability and lack of support have prevented widespread adoption of previous battery-free platforms."

That was the thinking behind Riotee, the open source hardware and software platform introduced by Nessie Circuits in early 2023. "Every year millions of new portable IoT devices are sold, and they are all powered by batteries. Regularly replacing millions of batteries is inconvenient, expensive, and bad for the environment," the pair said at the time. "We believe it's time for a responsible Internet of Things that leaves batteries behind in favor of renewable energy and sustainable energy storage."

Produced as part of a successful crowdfunding campaign, the Riotee platform is based on the Riotee Module — a compact system-on-module which includes a Nordic Semiconductor nRF52833 wireless microcontroller, some Texas Instruments non-volatile memory, an Ambiq real-time clock (RTC), and power hardware including a Maxim Integrated MAX20361 boost charger — all with a view to running the device, with its single Arm Cortex-M4F CPU at up to 64MHz, from energy harvesting systems with no intermediate battery.

Since the launch of the Riotee platform, the company has been gathering data on its use — and has used that to write a study, presented at the 22nd ACM Conference on Embedded Networked Sensor Systems (SenSys '24) in China this month. Perhaps the most interesting example provided: a battery-free device, based around a Riotee Board, Sensor Shield, Solar Shield, and Capacitor Shield, which hosts a TensorFlow Lite speech recognition model on a repeated one-second audio sample from a connected microphone to listen for a "hot" word — all under nothing but solar power from a nearby window.

"The battery-free device reliably distinguishes between 'yes,' 'no,' and silence/noise," the pair claim of their experimentation. "Preprocessing and feature extraction takes 251ms, while the actual inference takes 41.2ms. Under the given light conditions, a complete cycle from sound detection to result transmission takes 5.8s. The code of this battery-free application is available as an example in the Riotee SDK [Software Development Kit]."

Riotee Modules, Boards, and associated accessories are available to order on the company's Crowd Supply campaign page; hardware and software sources are provided on GitHub under the weakly reciprocal version of the CERN Open Hardware License Version 2 and the permissive MIT license respectively. The paper, meanwhile, is available on the ACM Digital Library under open-access terms.