Amazon’s new robot called Astro is designed to track the behavior of everyone in your home to help it perform its surveillance and helper duties, according to leaked internal development documents and video recordings of Astro software development meetings obtained by Motherboard. The system’s person recognition system is heavily flawed, according to two sources who worked on the project.
The documents, which largely use Astro’s internal codename “Vesta” for the device, give extensive insight into the robot’s design, Amazon’s philosophy, how the device tracks customer behavior as well as flow charts of how it determines who a “stranger” is and whether it should take any sort of “investigation activity” against them.
Videos by VICE
Subscribe to Motherboard’s CYBER podcast wherever you listen to podcasts.
Sentry
In its announcement video, Amazon says Astro is designed to give “peace of mind” to its owners. First and foremost, Astro is a surveillance device that tracks you and everyone who enters your home. When a user purchases the $999 robot, customers are asked to “enroll” their face and voice, as well as the faces and voices of anyone who’s likely to be in a home, so Astro can learn who is supposed to be there.
One of the internal documentation files presented in a development meeting and obtained by Motherboard describes how Astro patrols an owner’s home and tries to identify people it encounters. Other files refer to “Sentry,” the components and software that control the device’s security features. Sentry software includes integration with Ring cameras and Alexa Guard, Amazon’s home security service.
The meeting document spells out the process in a much blunter way than Amazon’s cutesy marketing suggests.
“Vesta slowly and intelligently patrols the home when unfamiliar person are around, moving from scan point to scan point (the best location and pose in any given space to look around) looking and listening for unusual activity,” one of the files reads. “Vesta moves to a predetermined scan point and pose to scan any given room, looking past and over obstacles in its way. Vesta completes one complete patrol when it completes scanning all the scan point on the floorplan.”
“Astro is terrible and will almost certainly throw itself down a flight of stairs if presented the opportunity.”
If the robot detects something it thinks is out of the ordinary, such as seeing a person it doesn’t recognize, or a sound like glass breaking or a fire alarm, it will investigate further, including following an unidentified person around the house, the file reads.
“Sentry is required to investigate any unrecognized person detected by it or Audio Event in certain set of conditions are met,” one file reads. “Sentry should first try to identify the person if they are not still unrecognized for as long as 30s [seconds]. When the person is identified as unknown or 30s passed, Sentry should start following the person until Sentry Mode is turned off.”
A flow chart presented during the meeting explains exactly what happens when Astro detects a “presence” and how it is designed for “investigating strangers.” If a user has disabled “stranger investigation,” the robot will ignore a stranger. If it’s set to “Sentry mode” or a patrol mode, it will either approach the stranger or follow them, and begin a series of “investigation activities,” which Amazon describes as “a series of actions Sentry takes to investigate audio or presence while recording.” Generally, if Astro begins an investigation, it will follow the stranger, record audio and video of them, and then automatically upload a recording the user can view later.
A user can tell the robot to start or stop a patrol either via a companion app or by saying a specific phrase out loud, such as “Echo, ask astro to start home monitoring,” or “Alex stop moving.”
The user can also put the robot into an “Away” setting, to signal that the user is not at home, and needs the robot to periodically patrol on its own. The user can then livestream and view what the robot sees to another device, such as their phone, while they’re not at home. Users can also start two-way video calls through the companion app and the robot.
“When vesta completes one complete patrol loop, a schedule for the next patrol is created, post one hour the time of current patrol activity,” one of the files reads.
A Flawed Robot
Developers who worked on Astro say the versions of the robot they worked on did not work well.
“Astro is terrible and will almost certainly throw itself down a flight of stairs if presented the opportunity. The person detection is unreliable at best, making the in-home security proposition laughable,” a source who worked on the project said. “The device feels fragile for something with an absurd cost. The mast has broken on several devices, locking itself in the extended or retracted position, and there’s no way to ship it to Amazon when that happens.”
“They’re also pushing it as an accessibility device but with the masts breaking and the possibility that at any given moment it’ll commit suicide on a flight of stairs, it’s, at best, absurdist nonsense and marketing and, at worst, potentially dangerous for anyone who’d actually rely on it for accessibility purposes,” the source said.
Another source who worked on the project mentioned privacy and navigation as chief concerns. “As for my personal opinions on the device, it’s a disaster that’s not ready for release,” they said. “They break themselves and will almost certainly fall down stairs in real world users’ homes. In addition it’s also (in my opinion) a privacy nightmare that is an indictment of our society and how we trade privacy for convenience with devices like Vesta.”
The source also corroborated that Astro’s facial recognition abilities perform poorly, which is concerning for a device designed mainly to follow people around and determine if they’re a stranger or not.
The robot will also closely integrate with Amazon’s Ring security system, including by responding to “Ring events” such as the aforementioned fire alarm, and automatically patrolling when the security system is set to away.
Kristy Schmidt, senior PR manager for devices and services at Amazon, told Motherboard in an email that regarding the person recognition, “in addition to consulting with several Amazon Scholars who specialize in computer vision, we also consulted with an external expert in algorithmic bias, Ayanna Howard, dean of the Ohio State University College of Engineering, to review the steps we took to enhance the fairness of this feature,” repeating what Amazon said in a blog post discussing the company’s visual ID feature.
“These characterizations of Astro’s performance, mast, and safety systems are simply inaccurate. Astro went through rigorous testing on both quality and safety, including tens of thousands of hours of testing with beta participants. This includes comprehensive testing on Astro’s advanced safety system, which is designed to avoid objects, detect stairs, and stop the device where and when necessary,” Schmidt added in an email.
“Amazon was thoughtful in the design, testing, and augmentation of their approaches as driven by data and feedback in order to minimize bias from their visual ID feature. They defined performance targets relevant to their products’ use cases, trained their models on a large volume of incredibly diverse data, and shared sufficient details of their methods with me in a genuine effort to ensure the feature not only works statistically well for all their customers, but that it also continues to get better over time on behalf of those customers,” Dr. Howard said, according to a quote provided by Schmidt.
Following Customers Everywhere
In order for Astro—which Amazon says will “improve customer’s lives in a way traditional consumer electronics can’t”—to work well, the robot must collect data about your behavior and your home.
“Vesta needs to learn customer behavior,” a document explaining Amazon’s “social robotics” philosophy states.
“The goal is to make Vesta an ‘intelligent robot,’ and allow some simple but magical interactions with people,” the social robotics document states. To do this, Astro needs to fully map a user’s home, creating a heat map of “choke points” and highly trafficked areas where the robot is likely to get stuck or “places where it will easily get hit by humans” such as hallways, doorways, and the kitchen.
Astro is supposed to learn over time, meaning that it must track what humans are doing, where they are going, and where they are likely to congregate.
“It’s, at best, absurdist nonsense and marketing and, at worst, potentially dangerous for anyone who’d actually rely on it for accessibility purposes.”
“The mobility in the device needs to work perfectly. I’m saying perfect here because if it’s not working perfect, we cannot expect social robotics to be welcome [in people’s homes],” a developer who worked on Vesta states in a meeting.
In order to be perfect, Astro must “appear ‘smart’ to users,” “needs to interact well with people, taking social norms into account,” needs to track and learn from customer behavior, and “needs to have some level of autonomy. It is available when needed, and it ‘wants’ to be around humans but not underfoot.”
Schmidt told Motherboard that “we designed Astro to handle a lot of the data processing on the device, including the images and raw sensor data it processes as it moves around your home. This helps Astro respond quickly to its environment. In addition, your visual ID is stored on the device, and Astro uses on-device processing to recognize you.”
Amazon’s programmers also mentioned in a meeting that they have determined the optimal distances Astro should interact with customers from. For example, during most times it should be “socially distant,” meaning 1 to 1.5 meters away from a user. However, “when a user is buying something from Amazon, the device knows the [screen] needs to be in a viewable state. 40 centimeters away” from the customer.
Jason Koebler and Jordan Pearson contributed reporting.
Update: This piece has been updated to include additional comment from Amazon.
