Training and testing on way more edge cases, for dramatically improved smoothness and safety.
Edge cases are the countless tricky or hazardous vehicle scenarios which are individually unlikely, but together make up all the risk on the road. If we can retrain AVs to contend with all the crazy things humans can throw at them, normal driving will become a walk in the park. Or as engineers put it “build for edge cases and get the center cases for free.”
Using an unprecedented scale and variety of data and patented AI, dRISK has assembled and mapped out out the whole space of edge cases into a knowledge graph of scenarios (like Google's knowledge graph of the internet, but in this case, of the real world).
Real driverless cars aren’t widespread yet because they aren’t safe, and they aren’t safe because they don’t understand the world the way humans do.
Modern autonomous vehicles can be tremendously sophisticated, but they are still basically elaborate demos. No matter how many billions of dollars are spent developing them, we will never have truly driverless cars until we solve this fundamental problem.
Advantages
Object detection re-trained with dRISK data can recognize edge cases 6x sooner and with 2x greater accuracy* than when trained on industry standard data, without a significant decrease in performance on easy "center cases."
dRISK is now releasing its edge case data to an exclusive group of users. Easy to consume and use, and rich with orders of magnitude more edge cases than any AV developer has previously had access to.
AI Engineer. Federico completed his MSc in Artificial Intelligence at the University of Edinburgh in 2021, with an emphasis on Deep Learning applied to Natural Language Processing and Computer Vision. His dissertation focused on developing an Unsupervised Synthetic Data Generation architecture with Unity, to generate synthetic data that closely resembles a set of real data, and can be used to train Computer Vision models. Prior to this, he earned an MSc in Mechanical and Industrial Engineering at Arts et Métiers ParisTech in France.
Automated page speed optimizations for fast site performance
