< BackHome/news/autonomous-cars-learn-from-humans

Let Me Do the Driving

autonomous vehicles

If there’s one thing the past 12 months have proven, it’s that self-driving cars are capable of making mistakes, sometimes leading to fatalities. Autonomous systems don’t always do well in, say, snowy weather or when street markings are faded out. Sometimes it’s that situation the car gets right 90 percent of the time, only to mess up the remaining 10 percent of the time due to anomalous situations (such as not pulling over when an ambulance with a siren is mistaken for another white passenger car).  New methods by Microsoft and MIT researchers aim to increase the awareness and improvement of these blind spots by enabling humans to correct autonomous driving systems in real time. In current testing taking place in virtual video game simulations,  human drivers manually handle the steering wheel, but the autonomous system checks itself simultaneously for any divergences and marks those down for future correction. Similarly, drivers are also able monitor self-driving systems in real time, taking over the wheel when necessary (kind of like it is now, but with the aim of real training involved). Learning from humans and synthetic data in and out of video games is nothing new in autonomous system development, and it might seem obvious, but until now this particular approach hasn’t been implemented in any real time way. Next stop: the real world. [Image: MIT News]

View Full Article >

sign up for our newsletter