Pete Warden’s Startup Puts AI in the Sensor

Article By : Sally Ward-Foxton

A founding father of tinyML co-founded Useful Sensors to help appliance manufacturers add AI capabilities to everyday objects with AI-enabled sensor modules.

Pete Warden, the former Google engineer widely seen as one of the founding fathers of the tinyML movement, recently quit Google and formed a startup to develop AI-enabled sensor modules. Useful Sensors hopes to bring AI capabilities to sensors for consumer electronics and home appliances.

TinyML refers to AI or machine learning (ML) running in resource-constrained environments, typically microcontrollers. Warden, formerly the technical lead on the TensorFlow Mobile team at Google, previously founded Jetpac, an early AI startup acquired by Google in 2014. He also published a textbook on tinyML.

By founding Useful Sensors, Warden intends to accelerate the addition of AI-enabled features to home appliances, including everything from light switches to TVs.

“I really wanted to have something that lets you look at a light switch and say “on,” and have the light go on,” Warden told EE Times. “That should just work! Or when I get up from my TV to make a cup of tea and I’ve got my hands full, I want the TV to pause. Or when I’m giving a slide presentation, I want to be able to swipe to advance to the next slide. These are all use cases for tinyML that we’ve been talking about for years.”

The Useful Sensors team discussing the company's AI-enabled sensor modules.
The Useful Sensors team. From left to right: Niranjan Yadla, Ali Zartash, Nat Jeffries, Manjunath Kudlur, and Pete Warden. (Source: Useful Sensors/Manjunath Kudlur)

Machine learning can help add this kind of intelligence to everyday objects in a way that doesn’t require huge compute, power consumption, or cost. Warden has been a little frustrated, however, with the rate of uptake of this technology by consumer electronics and appliance manufacturers.

Despite work by Warden’s team at Google to develop the open-source ML framework TensorFlow Lite for microcontrollers, plus Warden’s book and efforts from the community and TinyML Group on examples, courses, and conferences, uptake is still quite slow.

“Whenever I go to [appliance manufacturers], I tell them about all this wonderful free software that’s available for them to pick up and use, but usually at the end of it they say, ‘We barely have a software engineering team, we definitely do not have an ML team–can you just give us something that gives us a voice interface or wakes up our TV when someone sits down in front of it, and can you also give to us it for [a couple of dollars]?’,” Warden said.

With Useful Sensors, Warden aims to provide consumer electronics and appliance manufacturers with “something they can actually use, something that meets their requirements.” The AI-in-the-sensor approach echoes years of work on IoT smart sensors and sensor fusion devices.

“There’s a long tradition of heading in this direction,” Warden said. “We’re really trying to solve end-to-end problems, going the last mile to provide something that doesn’t require significant customization to be able to use. This is a solution that you can use off the shelf to solve a particular problem.”

The company’s first product is a person sensor: It’s a 20 x 20-mm board with a camera on the front and a microcontroller on the back. The board has two outputs: a single pin that goes high when a person is detected, plus an I2C interface for information, such as where people are in the camera frame, whether the person is looking at the device and basic facial recognition (enough to distinguish between family members using the same coffee maker, for example).

Useful Sensors' AI-enabled sensor
Useful Sensors’ first product is a 20 x 20-mm board with a camera on the front and a microcontroller on the back. (Source: Useful Sensor)

Useful Sensors is already talking to potential customers about how they can use this board.

So far, Warden said potential customers have been interested in applications like:

  • a fan that follows the user around the room.
  • a laptop that locks its screen when you’re not using it.
  • a surround sound system that knows where in the room people are seated.

Gesture control is another area of interest, Warden said, adding that most interested parties so far have been TV and laptop manufacturers rather than appliance manufacturers.

Dataset creation

Useful Sensors isn’t developing its own chip. Instead, the company is sticking with microcontrollers, at least for now.

The company sees its value add in dataset creation and model development, targeting companies whose core business doesn’t include building their own models and datasets.

Do customers not want to do ML development, or is it just that the early-stage software is an insurmountable barrier? Warden agrees early-stage software and a fragmented hardware landscape is part of the problem.

“Even if all those [issues] were fixed, you’d still have to learn how to create a data set to train an ML model, and that’s a very different skill set,” he said, adding that there is a lot of work and knowledge required to make high quality datasets, and for many, that is outside their core business.

Datasets for training ML models, even tinyML models, must be of the highest possible quality to ensure reliability. In the case of person detection, that means ensuring the dataset represents all kinds of people so that the model can be as reliable as possible across many different circumstances and contexts. To some extent, Useful Sensors intends to leverage community users to look for gaps it may not have spotted.

“This is a big part of why we’re trying to engage the maker community,” Warden said. “We want feedback from people out there to let us know if there are things we’re missing… we’re also trying to come up with ways of testing [the sensors] to see how well they work for different communities and different types of people through third-party testing.”

Security and privacy

Warden is also keenly aware that adding cameras to household appliances poses security and privacy questions. Warden hopes to have the company’s person sensor certified by third parties to ensure hackers can’t access its camera. The module’s I2C interface carries only metadata about the scene (not full frame images), and there’s no network connection on the module.

“TVs and laptops are in people’s bedrooms. That’s a massive responsibility,” he said. “We really believe that this is going to be a better privacy solution than having something that’s integrated into the rest of the system.”

Different types of sensors are also on Useful Sensors’ roadmap, with the most likely next product being a speech-recognition sensor, again using microcontrollers rather than more specialized chips.

Useful Sensors raised a $5 million seed round and currently employs six people, three from Google. Warden is co-founder and CEO, with co-founder and CTO Manjunath Kudlur also joining from Google’s Tensor Flow team (via Cerebras).

Useful Sensors’ person sensor board is available now.


This article was originally published on EE Times.

Sally Ward-Foxton covers AI technology and related issues for and all aspects of the European industry for EETimes Europe magazine. Sally has spent more than 15 years writing about the electronics industry from London, UK. She has written for Electronic Design, ECN, Electronic Specifier: Design, Components in Electronics, and many more. She holds a Masters’ degree in Electrical and Electronic Engineering from the University of Cambridge.


Subscribe to Newsletter

Leave a comment