When it comes to joining the self-driving car race, it seems it’s all fun and games for Amazon.
The internet giant has officially entered the fray—sort of—with the help of an unlikely source: a toy car, no more than 1/18th the size of its automotive cousins.
Called the AWS DeepRacer, the tiny vehicle represents big possibilities for the company. Amazon plans to make the product available this March.
It’s the remote-controlled car of the future, fully capable of driving itself—that is, once it gets a little assistance from humans.
For just $399 (or $249 presale), purchasers will receive the fully-programmable toy stocked with an HD camera, dual-core Intel processor, and other hardware needed to get itself around. But humans are on the hook for teaching it how to drive, which is done using cloud-based platforms and simulators also provided by Amazon.
To sweeten the deal, the internet purveyor has simultaneously launched the “world’s first global self-driving racing league,” where DeepRacer owners can watch their projects battle it out on the track in real life.
Still, the enterprising company may be getting much more than a few hundred dollars and a bit of entertainment from the miniature rides.
The self-driving toy cars utilize a software called “reinforcement learning,” which has proven to be an extraordinarily successful tool for developing AI. And while Amazon has said it hopes that the toy’s release will help familiarize more users with that type of programming, the company has remained tight-lipped on what, exactly, will—or could—be done with all the data the product generates.
If at First You Don’t Succeed…
The DeepRacer gains its eventual autonomy through a process that may look very familiar to its human counterparts: trial and error.
Similar to other deep learning methods being deployed in increasingly smart machines, reinforcement learning (RL) allows a computer to rely not on a pre-programmed route or even a selection of pre-programed choices, but on its own actual experience.
Once a computer has seen a scenario come up several times, it can then begin checking its own memory banks and evaluating previous outcomes to help determine the best route forward.
The method works by repeatedly exposing the machines to certain situations, allowing them to make mistakes along the way in order to learn. Once a computer has seen a scenario come up several times, it can then begin checking its own memory banks and evaluating previous outcomes to help determine the best route forward.
The algorithm flourishes in an interactive environment, but the computers will learn just as well from running through simulated situations, which is where Amazon comes in.
The crafty company also hosts several machine learning programs, including Amazon SageMaker, which DeepRacer owners can utilize to get their cars up to speed more quickly.
But the ultimate carrot for RL-taught machines is working toward a pre-determined goal or within a reward-based system. This is the same method Amazon is using on purchasers of the product, with the introduction of the robo-racing league.
Due to launch in 2019 at the company’s series of AWS Global Summits, the races will welcome anyone working on their mini self-driving cars, letting the autos duke it out in the ultimate interactive environment.
But Amazon will truly reap the rewards.
For What It’s Worth
Reinforcement learning has already proven to be a proficient development tool for artificial intelligence—with nearly unending potential for profits.
The software design first came to prominence last year, when it helped a computer defeat the reigning world champion in the complex board game Go. The victorious machine, called AlphaGo, was built by one of Amazon’s biggest corporate rivals, Google parent company Alphabet.
The explosion of development that’s bound to occur once Amazon releases its race car to the masses will likely usher in the foundation for solutions to any number of other problems.
But the algorithm has since been applied to any number of uses more consequential than winning a game.
Its ability to teach computers how to adapt to a changing environment makes it especially helpful in situations that require carefully calibrated responses, such as when machinery must interact with unpredictable weather patterns or the power grid. General Electric has already utilized the software to help improve image-processing models in MRI machines.
Still, the explosion of development that’s bound to occur once Amazon releases its race car to the masses will likely usher in the foundation for solutions to any number of other problems—and that the codes will be developed completely on Amazon-owned properties gives the company a compelling case to claim ownership over any breakthrough material.
Amazon isn’t the only one employing this crowd-sourced philosophy. Fellow Seattle heavyweight Microsoft has also released an open-source simulation for coders to apply reinforcement learning on cars and drones.
But if Amazon has the luck and inclination, it may be able to exploit the slipstream created by those pushing RL forward, to draft its way into the lead of the self-driving car race.
