Simulations won’t make GM Cruise a truly autonomous car

gm cruise
Image credit: GM Cruise

GM Cruise is one of the leaders in autonomous driving, but I suspect that its use of simulation as opposed to the real world will prevent its performance from getting to a level where it can be truly called autonomous.

GM Cruise held an event last week where it discussed its autonomous driving technology in detail with a focus on two areas in particular.

First, custom silicon: Like Tesla and NVIDIA, Cruise has designed its own silicon that it intends to include in its vehicles that will execute the autonomous driving. This makes sense because one of the biggest hurdles to the adoption of autonomous driving (excluding technology) will be the cost of deploying it into vehicles. Thanks to the cost of lidar and compute, autonomous driving is both very expensive to deploy as well as a significant drain on resources. In EVs, this is obviously a problem as it will sap the vehicle’s range, which remains one of the key differentiators from one proposition to the next. There is not that much that Cruise can do about lidar, but following Tesla down the custom silicon route makes a lot of sense.

Second, simulation: GM Cruise is increasingly using simulation rather than real-world driving to iron out the kinks in its autonomous driving software. This is done by mapping a real-world city and then recreating that city within a simulation so that the software’s ability to navigate that city can be tested and refined.

I consider the biggest problem with autonomous driving not to be the decision-making capacity of the system, but the perception of the environment upon which those decisions are based. Hence, the biggest problem to solve is the interpretation of the real world accurately such that the machine can make good decisions.

Almost all of the horrible mistakes I have seen Tesla and everyone else make are not based on decision making but on the incorrect perception of reality. Hence, when one recreates reality in simulation, the task becomes easier as much of the perception that the system must do has already been done for it.

Furthermore, simulations are limited to the rules upon which they have been based and therefore may not be an accurate representation of reality.

This is why I am pretty sceptical about the ability of simulation to produce an algorithm that can drive reliably and safely in the real world. There are uses for simulation in the training of algorithms, but I think it will struggle to deal with the endless corner cases that keep tripping all of these systems up.

I have long believed that the solution that is most likely to work will be one that combines deep learning systems and software. This is because deep learning can learn but not reason, and software can reason but not learn. Hence, combining the two should be complementary and help offset each other’s weaknesses.

Despite both Cruise and Waymo receiving licences to operate driverless vehicles in California from the DMV, they also need a licence from the California Public Utilities Commission (CPUC) and the time scale on this is completely unknown.

Hence, I think that we are still pretty far away from true driverless vehicles although there will continue to be vanity projects launched here and there. I do not see anything that would lead me to change my estimate of 2028 before autonomous driving is properly commercialized meaning that many start-ups are going to run out of money. There has already been considerable consolidation and I expect more to come.

Be the first to comment

What do you think?

This site uses Akismet to reduce spam. Learn how your comment data is processed.