Why have Self-Driving Cars failed? In our recent Twitter & Blog post written we argue that this is mainly due to lack of real out-of-distribution training & testing data for advanced computer vision models in the perception engine for Self-Driving Cars.
Indeed, the current Self-Driving Car Market ($207B) is passing through a global inflection point where there are many companies, pushing through, struggling and also going bankrupt. What has happened in the latest years for this market to get to this point? At Artificio, we have identified that at the center of this crisis is the over-promise of technology at the three stages of self-driving autonomy: perception, navigation & control. In particular in perception (computer vision + LIDAR), the autonomous driving industry has suffered tremendously given the flagrant errors that have happened when these systems encounter out-of-distribution objects & adversarial stimuli never seen before at training time. And although this is very difficult to believe given the recent advances of Self-Driving cars locally working as Taxi's in San Francisco, there are still many failures.
Our start-up thesis is that we will need to train and test self-driving systems on a larger variety of out-of-distribution data (also known as corner-case data) to both: 1) prevent accidents in production, by correctly recognizing an averse situation and it's objects; and 2) to stress-test these perceptual systems pre-production (in simulation on the Cloud) to know if they can handle critical and unexpected real-world scenarios.
In other words, companies wanting to generate real out-of-distribution (crazy driving) data from San Francisco, can not do this in practice if they only have access to in-distribution (regular driving) training data. This creates an opportunity and un-fair advantage for us to seize where we capitalize on the commercialization of real out-of-distribution data that can be crowd-sourced from emerging economies that share certain traits with their developed counter-parts.
Driving in places like Peru, Colombia, Brazil and Mexico are chaotic edge-case scenarios of driving in United States, Canada or China. Algeria, Egypt, Morocco, and other countries in Northern Africa and Eastern Europe may provide value in self-driving car companies in France and Germany. In addition, the overflow of traffic, and visual chaos of left-lane driving nations like India, Sri Lanka, Thailand and Malaysia provide the perfect stress-tests for the United Kingdom, Singapore & Japan. While many GenAI companies are chasing the AI hype and are trying to synthesize new data through a ''Digital Twin'', our angle is to exploit and find the ''Analog Twin'' city of places like San Francisco (eg. Bogota, Colombia), and get as much real out-of-distribution data that will help Self-Driving cars currently in San Francisco, a first step that will help us bring Autonomous Driving to the rest of the world.
Follow our latest news and Progress on X/Twitter : @Artificio_Org
