Robot vacuums are learning to avoid dog poop, but that’s not all they can see

In the lab, robots are engineered for mind-blowing displays of perception and agility, like Boston Dynamics’ parkour performing bot and the humanoid shooting hips at the Olympics.

But in the home, this technology is now deployed for the mundane task of spotting dog poop — a feature with expansive potential privacy costs.

“Roomba j7+,” the latest version of iRobot’s popular home vacuum, claims to give customers “even more control over their clean,” with a camera that can identify and avoid pet droppings. Instead of smearing it all over the floor, the device will gracefully avoid the poop and even snap a picture and text it to your phone if you’re out, the company says.

The $849 vacuum, released last week, relies on an AI-boosted brain and camera system to identify objects on the floor in real-time. It’s designed to be a “thoughtful, collaborative cleaning partner” suited for people who want tech to serve them better, iRobot says in a news release.

For the company that pioneered robot vacuums, it represents a significant upgrade.

“Smart home products often fail to live up to consumer expectations when they lack context of the home, cannot learn independently and require complex programming for basic functionality,” said iRobot CEO Colin Angle. “We think we can deliver a [smart] home [product] that actually meets your need.”

But there’s a catch: Accessing these features requires you share sensitive household data.

The company touts that the gadget gets smarter every time you use it, a feature made possible by an AI and vision platform, trained on tens of thousands of photos taken inside iRobot employees’ homes, the company says. The photos help the vacuum make sense of what it is confronting as it scoots around your floor, and AI allows it to personalize how and when it cleans.

To maneuver around pet poop, the luxury vacuum uses a set of sensors to navigate around a room and map your furniture layout. If an object lies in its path that it suspects is a chord or dog poop, it’ll take images of the objects and send them to a smartphone app.

iRobot built over 100 physical models of pet droppings and trained algorithms on over 100,000 images to get the device to avoid crap, the company said in a video posted on Twitter.

If the gadget continuously encounters cords in a specific part of the room, it might mark off that area as a potential “keep out zone” for you to confirm on the app. It may even learn when and where you typically like for it to clean and figure out more opportune times to roam around based on when you’re away from home, Angle said.

It’s not just iRobot rolling out advanced vacs with more powerful vision and decision-making capabilities. In January, Samsung revealed a “JetBot90 AI+” with object recognition backed by Intel’s AI software. The $1,300 vacuum doubles as a home security camera, allowing people to livestream their rooms from robot eye level. The device is also supposed to avoid dog droppings and cables, the company says.

But experts say AI-powered home devices raise privacy considerations, in part because of the intimacy of the data they collect.

“People are used to thinking about whether Alexa is listening in on their house or what the Ring doorbell is capturing outside, but they might not realize that the existence of a camera on their vacuum could present those same types of concerns,” said Tom Williams, assistant professor of computer science and the director of the Mines Interactive Robotics Research Lab at the Colorado School of Mines.

While audio-based home assistants are common, “once you move video into the home, it results in a bunch of new things being captured that audio alone wouldn’t be able to do,” he said.

Since 2018, Roombas have used sensors and a low-resolution camera, pointed at the ceiling, to map your home so the vacuum can move swiftly. But with a front-facing camera, the latest model can see anything left on the floor and more easily avoid obstacles.

For customer privacy, the company says the vacuum only recognizes three specific objects: chords, pet droppings and its charging base. Angle says the software automatically shuts off the camera if it detects a human or photo of a human within view and finds a human-free angle to capture. He also says the firm will never sell user data: Images taken by the Roomba are stored on the device, not on the cloud, unless you agree to send them to the smartphone app or to iRobot.

If you opt in to sharing them with the company, those images are encrypted and sent to iRobot, where staffers unencrypt them and compare them to other images in the company’s database and use them to further train its robot vacuums, said Michelle Gattuso, product management vice president of portfolio and software experiences at iRobot.

Still, there are concerns about what happens should Roomba ever decide to share this data or is unwittingly hacked.

“In theory, they could sell granular information to a Redfin or Zillow that’s trying to get an understanding of the interior of a home that’s not for sale yet,” said Ryan Calo, a law professor at the University of Washington researching policy and emerging technology.

iRobot says all images are deleted after 30 days. Since images taken of objects are encrypted, if someone were to hack the robot, they wouldn’t glean much, the company claims.

“Even if we were hacked and someone got the map of your home, they would know that you have a kitchen. They would know that you have something in that kitchen called the table. They would not be able to see what your kitchen looked like,” Angle says. “What we store is very abstract. It’s a line drawing.”

Outside experts say people should remember that wherever there is data, there exists a chance for someone to see it and interpret what it means.

“The attribute of data is that it is observable because it exists,” said Karl Rauscher, founder of the Association of Cloud Robot Operators, a robotics consulting firm. “If it’s only observable by the person who has the encryption key, then any of these things could be compromised through permutations, or by simply accessing the key.”