UC Berkeley's BADGR Robot Uses AI to Learn How to Navigate
The autonomous robot can learn to navigate in real-world environments without any simulation or human supervision.
Most autonomous robots use geometry-based systems to navigate collision-free from one point to another; it's a form of the predictive path approach that allows vehicles to avoid obstacles while heading to a predetermined destination. The problem with that is the vehicle "thinks" it can't travel to that point via different terrain other than what it's programmed to do. For example, if an autonomous robot needs to retrieve mail, it will travel down the driveway to get to the mailbox, rather than rolling across the grass to reach its goal.
To solve that problem, a team from UC Berkeley designed the BADGR (Berkeley Autonomous Driving Ground Robot), an end-to-end autonomous robot that uses AI to gather data in real world environments without any simulation or human supervision to get to its destination. In this instance, it eschews the predictive path algorithm in favor of learning from experience.
In a recent paper, the engineers state: "BADGR can navigate in real-world urban and off-road environments with geometrically distracting obstacles. It can also incorporate terrain preferences, generalize to novel environments, and continue to improve autonomously by gathering more data."
The team outfitted the robot with a neural network system that gathers real-time data via camera and uses that information combined with future planned actions. It will then look for obstacles or possible collision points, and determine the best avenue of approach to a destination, meaning it will decide if it rolls across the lawn or takes the driveway with a bunch of bikes on it to fetch the mail.
Powering the BADGR is an NVIDIA Jetson TX2 that processes information gathered from the camera, along with a six degrees-of-freedom IMU sensor, GPS, and a 2D LIDAR sensor, which it uses to plan the desired path and then learn how to get to a destination more efficiently. While the platform is a great step toward a fully autonomous self-improving navigation system, it still has its problems, including how to safely gather data when in new environments. Does it take the reliable dirt path on the side of a mountain, or does it dive off of a cliff? I guess we will find out in future revisions.