Microsoft Garage’s Project Road Runner looks to solve autonomous driving problems

Kareem Anderson

Looking for more info on AI, Bing Chat, Chat GPT, or Microsoft's Copilots? Check out our AI / Copilot page for the latest builds from all the channels, information on the program, links, and more!

How do you teach a car to drive? Throw pictures at it, apparently.

Perhaps, that was a crude way to describe Microsoft’s foray into developing autonomous driving algorithms for a future where driverless cars roam the streets in mass.

Recently, the Microsoft Garage team released a post on its self-titled blog detailing a new deep learning project by the name of Project Road Runner in which the team is using photo-realistic simulations to help train driverless cars.

As with most of the company’s Garage efforts, the team wasn’t officially backed by the Redmond based software giant and had to seek resources outside of the multi-billion dollar treasure chest of Microsoft. That resulting innovation spawned from the creative testing methods of the Garage team resulted in the team seeking simulations versus using actual vehicles encased in expensive hardware and sensors.

Furthermore, the Garage team sought an open-sourced simulation platform AirSim, that, at the time, was being used for drone simulation, to help them conduct their shoestring budgeted research. Aside from its cast landscape footage, AirSim is also credited with particularly useful photo-realistic environments built atop the Unreal gaming engine. The net result of this usage by the Garage team helped spur a symbiotic partnership that birthed AirSim’s autonomous driving expansion in November 2017. Now, the software to help train autonomous vehicles is free to anyone outside of Microsoft.

I’m sure both Google and Uber have already figured out the limitations Project Roadrunner’s approach, but it suffices to say that, with simulation driving, the Garage team was free to gather weekly Petabytes worth of data on reinforced learning by allowing their vehicles to freely crash into things, without fear of accidental death or lawsuits.

“This is pretty big. With reinforcement learning, you eliminate the need for collecting and storing petabytes of data every week. You can let your car crash into things and learn on its own. Reinforcement Learning has a lot of potential in the AD space, and of course it is impossible to do it in the real world, which is just another reason why I think simulation is going to be the backbone of the AD industry going forward. With our latest tutorial, you can start doing this at scale by distributing the learning job across multiple VMs on the cloud.”

Despite not having a fully promoted play in the war for infotainment unit space or a well a well-publicized footprint in autonomous vehicles, Microsoft has slowly been crafting a narrative as being the connective tissue that binds software to hardware via the cloud and Project Roadrunner is an exemplary showcase of that idealism. Using open source software to bypass current hardware limitations and powered by the vast opportunities presented with cloud computing, Microsoft employees could be creating the foundation for autonomous vehicles. A drivers ed of sorts for driverless cars that could be implemented by car manufacturers for years to come.