Autonomous Levels 4 and 5 will never be reached without Simulation vs Public Shadow Driving
Every AV or SDC maker using public shadow drivers for AI will have to drive each vehicle type ONE TRILLION miles at an expense of at least $300B and will be putting those public shadow drivers and the public at risk. A risk that will result in injuries and fatalities as the scenarios progress from the currently benign to progressively more complicated and dangerous scenarios. Scenarios that are the most dangerous and complicated combinations of traffic patterns, bad road conditions, bad weather and a variety of moving vehicles and entity complexities. Even sensor or vehicle systems degradation.
The reason I am saying L4 and L5 will never be reached using this method is that most organizations cannot afford this path for various reasons. And the time will come when governments and litigation shut the process down.
Regarding the cost. One trillion public shadow driving miles in 10 years is 228k vehicles driving 24x7. (Since most vehicles cannot take 400k miles a year you will wind up using more than that 228k). Those driving 24x7 takes 3 drivers a day that is 684k drivers. (Drivers who have to be skilled enough to get every action right or AI doesn’t learn the right thing). My very conservative $300B estimate is for the vehicles, gas sensors and drivers only (At a rate of $20k for the vehicle and sensors). Beyond that cost will be the cost of litigation for the accidents, injuries and loss of life that will occur. You cannot drive and redrive dangerous complicated scenarios over and over, thousands or more times, and not have accidents. Especially for scenarios that are meant to learn what to do in actual accident scenarios. Are you going to have shadow drivers drive accident scenarios in bad weather, with bad road conditions with dozens of other vehicles, pedestrians etc around? And keep driving billions of miles restumbling on them to train AI? Just counting the known or anticipated accent scenarios you would causes thousands, hundreds of thousands or more accidents till you got it right. (And that doesn’t count AI getting it right but the other drivers overcoming that and causing an unavoidable accident). That will be thousands in not tens or hundreds of thousands of injuries and loss of life and property.
Now let’s factor in the litigation and government intervention. As it is no children have died in a public shadow driving accident to date. Even in benign conditions. (The current non-complicated or dangerous situations. Where the streets are well lined lines, mapped and learned and there are good weather and street conditions without a lot of complexity. That alone could shut the whole thing down for quite a while starting tomorrow). Let’s say it doesn’t happen and some time goes by until the dangerous scenarios are run. There is absolutely no way various levels of governments, insurance companies, lawyers and individuals let you turn public roads into accident scenario beta test or Guinea pig sites. When this happens everything will come to a grinding halt leaving most of the complicated, dangerous or accident scenarios unlearned. That will stop L4 and L5 progress. (Tesla, comm.ai, PolySync etc are on this path now with the public not paid drivers as those beta testers).
So what is the answer. . . Simulation for AI data gathering, engineering and testing. Augmented with test tracks and other sources where simulation cannot actually meet the burden and to ensure the accuracy of the simulation.
Let me address the first thing folks usually say at this point. That simulation is not up to this. YES it is. Why don’t most folks know that? Because most of them come from Commercial IT or even the automakers and they do not have exposure to nor experience in simulation in the aerospace industry. Which has had most of the capabilities needed for 20 years. Is this more complicated than that? Yes. But the technology is there. Beyond that I believe that some of the current simulation products are not that far away. The problem is that part of the industry is disjointed. Not everyone knows what is available, what the capability gaps are or how to close them. That is why I am proposing an international association and trade study exhibit be created. (We have recently determined we want to add test tracks and all non-public AI and testing entities to the association.)
For much more detail on all of this as well as references for information I cited please see the articles below
Who will get to Autonomous Level 5 First and Why.
Stop relying on AI to make Autonomous Vehicles — You are wasting time and risking lives
Autonomous Vehicle and Mobility Simulation Association and Trade Study