One of the biggest obstacles to full autonomous driving the long tail of unusual events that can occur on the road. How can autonomous vehicles learn to handle such rare events? It is relatively easy to collect large amounts of unlabeled data by placing sensors on a fleet of vehicles and driving around the world. The main bottleneck is accurately labeling such vast amounts of data; even with the best labeling tools, data can be collected much faster than it can be labeled. We propose to overcome this challenge with self-supervised learning, which can learn from on unlabeled data. We will present two of our recent work on self-supervised learning for autonomous driving: self-supervised scene flow and self-supervised data association. These methods will enable us to improve performance by training on large amounts of unlabeled data. Finally, we present an algorithm that uses a low-cost active depth sensor for object detection, which can enable safe and affordable autonomous vehicles.
Back to Workshop I: Individual Vehicle Autonomy: Perception and Control