This is how Amazon is training its Scout delivery robots

When engineers train AI agents to play video games, they often do so at an accelerated rate, having the bot go through hours of human-speed gameplay in a matter of minutes to stress-test its capabilities. Unfortunately, no such shortcut exists when developing AI that controls machines in the real world. You can’t fast-forward reality.

Well, you can if you simulate it, which is what Amazon is doing to test its new Scout delivery robot. Speaking to The Verge, Scout VP Sean Scott explains that the company has created detailed virtual maps of American suburbia to accelerate development of the bot. It’s collected 3D data, real-life textures, and modeled the sidewalk down to the storm drains.

“We can run thousands of deliveries in simulation overnight versus taking a bot outside in the real world,” Scott says. “The bot doesn’t actually know it’s in a simulation. It thinks it’s in the real world, which is pretty cool.”


Think of it like The Matrix but for delivery robots.

Scott adds that he doesn’t know of any other company “talking about this level of fidelity at this scale for this type of training,” and that Amazon’s other training apparatus include an indoor robot park, and special rigs to test the resilience of the bot’s wheels.

Amazon wheel testing rig for its Scout robot.
Image: Amazon

Being able to speed up development is important for Amazon, not least of all because it’s a relatively late entrant in the delivery robot game. The company only unveiled its six-wheeled, cooler-sized robot back in January, but a huge range of startups have been building robots in this space for years, testing them in small neighborhoods and closed environments likes offices and campuses.

Of course, if any company is going to make this tech a reality, it’s probably Amazon. Founder and CEO Jeff Bezos boasts of his company’s ability to pour huge financial resources into projects like this, and Amazon loves to experiment with weird delivery options, from drivers that leave packages inside your home to transforming delivery drones.

Scott says the focus in Scout’s development is just the opposite: making sure the robots don’t come across as unusual but as a natural part of the environment. That’s accomplished partly through industrial design, he says, with Scout’s friendly curves, toy-like rubber tires, and chirpy blue paint job giving it a distinctly nonthreatening vibe. “We want the robot to fade into the background,” he says. “We call it ‘design for boring.’”

Another example of ‘design for boring’? Amazon’s training park for AI drivers.
Image: Amazon

It also means modeling Scout’s physical behavior. Like all delivery robots, this is a machine that has to work alongside humans, navigating in a way that doesn’t irritate or endanger pedestrians. That means making way for children, pets, prams, and the elderly, but not letting this deference turn into indecisiveness.

It’s a tricky balance to strike, as self-driving cars have demonstrated. Studies show that pedestrians, drivers, and cyclists are likely to take advantage of any overly cautious software design, bullying the autonomous vehicles off the road.

Amazon is confident that it’s on the right path. As an example, Scott proudly shows off two videos of Scout encountering the same neighborhood dog. In the first, the dog is alert and inquisitive, wary of this new arrival in its territory. In the second, it barely pays attention to the bot. “What better way to gauge response than from a pet like this,” he says.

Some rival delivery bot companies like Kiwi Campus have been testing their machines for years.

Right now, Amazon’s operations are very limited compared to its rivals. It has just a handful of delivery robots operating in Snohomish County in Washington, and each has to be accompanied by human overseers to make sure the robot doesn’t get into trouble.

The company won’t say where or when it plans to expand these tests, but Scott says its simulations will help it move into new territories, allowing engineers to test the bot’s autonomous driving skills in the virtual world before the rubber hits unfamiliar roads. “The next step is to capture other geographies [and] start testing in simulation. And then, eventually, we’ll be delivering around the world.”

This is big talk, but it’s clear that Amazon — like other companies in this area — still has to work out a lot of kinks before delivery robots become commonplace. Beyond issues of reliability and safety, there are practical matters, like simply getting the package to the customer.

Amazon’s Scout robot can’t climb steps, so for the moment the company’s helpers have to take deliveries out of the robot and hand them over on the doorstep. (Some companies are considering legged robots to make this process easier.) There are also possible regulatory challenges, with some cities like San Francisco branding delivery robots a nuisance and banning them from the sidewalk completely.

Scott says the best way to prove that delivery robots can work is to just make delivery robots that work — starting with the virtual world and then the sidewalk. “We have a ladder to the Moon … but we’ve only made it to the first rung of the ladder,” he says. “We’re learning lots, and we’re just getting started.”