Images captured by iRobot’s Roomba j7 robotic vacuum cleaner were sent to Scale AI, a startup that tags audio, photo and video for artificial intelligence training. Later, 15 photos (including a picture of a woman sitting on a toilet with her shorts down) somehow ended up in closed groups on some social networks.

Other photos show rooms of houses, people and pets. Furniture, decor, and objects high on walls and ceilings are accompanied by labels such as “TV,” “plant,” and “lamp.”

iRobot – the world’s largest supplier of robotic vacuum cleaners, which Amazon wants to acquire for $1.7 billion – has confirmed that these images were taken by its Roomba models in 2020. All of them came from “special devices not intended for the average user”, but were transferred to testers and employees who officially agreed to send their own data through robots.

Course

FINANCIAL MANAGER

Become a professional financial manager and earn from $500 in 2 months.

REGISTER!

According to iRobot, the devices were marked with a bright green sticker that said “Video recording in progress,” and testers could remove anything they deemed sensitive.

The 15 “leaked images” are just a slice of the massive data ecosystem. iRobot says it has shared more than 2 million images with Scale AI and an unknown number with other data annotation platforms.

James Baussmann, a spokesman for iRobot, said in an email that the company “has taken precautions to ensure that personal data is handled securely and in accordance with applicable law,” and noted that the images were “distributed in violation of the nondisclosure agreement between iRobot and image annotation service provider”. iRobot CEO Colin Engle later said that “iRobot is terminating its relationship with the service provider that leaked the image.”

However, at the end of the day, this set of images is about more than the actions of any single company. He talks about the common practice of sharing potentially sensitive data to train algorithms, and the amazing journey a single image can take — in this case from homes in North America, Europe and Asia to iRobot servers in Massachusetts, from there to Scale AI in San -Francisco, and finally to Scale’s contracted data workers around the world (including, in this case, workers from Venezuela who posted the images in private groups on Facebook, Discord, and elsewhere).

Evolution of robot vacuum cleaners

Robot vacuum cleaners weren’t always so smart. The very first model, from the Swedish company Electrolux Trilobite, entered the market in 2001. It used ultrasonic sensors to determine the location of walls and additional impact or break sensors to avoid running into objects or falling down stairs. The sensors did not always work correctly and the cleaning was far from ideal.

In 2002, iRobot released the first generation of Roomba, which relies on similar basic shock and rotation sensors. The model was much cheaper than its competitor, and became the first commercially successful robot vacuum cleaner.

Base models today operate largely the same, while mid-range vacuums get the best sensors and other navigation methods, such as simultaneous localization and mapping, to map out the best paths for picking. High-end devices have begun using computer vision, including iRobot (which has sold over 40 million units since 2002), Ecovacs, Roborock, Samsung, LG and Dyson.

“Computer vision allows work to see all the wealth of the surrounding world. This allows iRobot devices to avoid cords on the floor or to sense what’s ahead, such as a sofa,” says Chris Jones, iRobot’s Chief Technology Officer.

Where do robots get data from?

For computer vision in robot vacuums to really work as intended, manufacturers need to train it on high-quality data sets. MIT Technology Review spoke to or inquired with 12 companies that sell robotic vacuum cleaners and found that they respond differently to the task of collecting training data.

In the case of iRobot more than 95% of the image data set comes from real-life homes inhabited by either iRobot employees or volunteers, employed by third-party data providers. The company also invites regular consumers to provide training data through its algorithm improvement program.

Roborock said it either “produces images in its own labs” or “works with third-party vendors in China who are specifically asked to capture and provide images of objects.” Meanwhile, Dyson, which sells two models of high-end robot vacuums, said it was collecting data from “testers in the research and development department” and from artificial intelligence-generated training databases.

Most companies have said outright that they do not use customer data for their machine learning algorithms. Samsung, LG and Bosch did not respond to requests for comment, while Ecovacs calls the source of its training data “confidential”.

Hints about other collection methods were provided by hacker Giese, who studied Dreame robot vacuum cleaners, for a new Chinese company. He found a folder labeled “AI server” in the models, as well as image upload functions.

“Companies often say that ‘camera data is never sent to the cloud’, but when I got access to the device I was able to prove that this is not true. Even if they really haven’t uploaded any photos, the feature is available,” says Giese.

Data annotation centers

The market value of data instruction service is projected to reach $13.3 billion by 2030. This field has largely evolved to meet the huge need for labeled data to train algorithms used in self-driving cars.

Among the many startups that have appeared in the last decade, Scale AI has become the market leader. Founded in 2016, the company built a business model by contracting remote workers in poor countries based on tasks on Retasks, its own crowdsourcing platform.

iRobot confirmed that the 15 images posted on social media came from its devices, providing a spreadsheet listing the specific dates they were created (June to November 2020) and the countries they originated from (US, Japan, France, Germany and Spain). . ), as well as the serial numbers of the devices that created the images and notes indicating that the consent form was signed by the user of each device.

iRobot claims sharing the images in social media groups violates Scale’s agreements with it, and Scale says the contract workers who shared the images violated their own agreements.

Companies mostly try to avoid photos with naked people, and generally hide their faces in any pictures. In the case of the woman on the toilet, there was a black circle on her face. But in other images with people, their faces were not hidden.

Perhaps you agreed to this

iRobot’s privacy policy states that it only collects audiovisual data if a person shares images through the mobile app.

LG’s privacy policy for Hom-Bot Turbo+ with camera and artificial intelligence explains that its app collects data, including “audio, electronic, visual or similar information such as profile pictures, voice recordings and videos.” And the privacy policy for Samsung’s Jet Bot AI+ Robot Vacuum and Powerbot R7070 with cameras will collect “information stored on the device, such as photos, contacts, text logs, touch interactions, settings, and voice command recordings.”

Roborock’s privacy policy does not mention audiovisual data, although company officials told MIT Technology Review that consumers in China can share it.

Helen Greiner, co-founder of iRobot, emphasizes that by collecting all this data, the companies are not trying to violate the privacy of their customers.

“They’re just trying to make better products — or, in the case of iRobot, make cleaning better.”

Still, even the best intentions of companies like iRobot clearly leave privacy gaps.

What’s next?

The appetite for data will only grow in the coming years. Vacuum cleaners are only a tiny part of the smart devices that spread in our lives. That big mission is seen in the fact that Scale’s data annotators marked not only objects on the floor to avoid (a feature touted by iRobot), but objects such as a closet, kitchen table, and shelf.

Robotic vacuum cleaner companies are already investing in other features and devices that will bring us closer to a robotic future. The latest Roomba can be controlled by voice through Nest and Alexa and they recognize more than 80 different objects in the house. Meanwhile, Ecovacs’ Deebot X1 robot vacuum has integrated its own voice assistance, and Samsung is one of several companies developing “robot companions.” Miele, which sells the RX2 Scout Home Vision, has focused on a smart oven with a camera.

And if Amazon’s $1.7 billion acquisition of iRobot goes through (currently pending FTC approval), the Roomba will likely become even more integrated into Amazon’s smart home vision. And then we should be more careful about the technology in our home, or at least read the privacy policy more carefully.

Source: MIT Technology Review