Regarding Cruise and GM.
TL;DR - For this thread and Tesla investors, Cruise poses no threat to Tesla, just as other EV car makers, in that their success will demonstrate that with AI and sensors, cars can drive without supervision by humans. There is simply market share for everyone at this point...
Longer version...
Cruise is progressing (which is great), they have moved the human driver from the driver seat to a remote location connected by a real-time feed as a supervisor with the ability to hit a big red dedicated stop button just like a human driver would have the ability to press the mechanical brake pedal. And if the real-time feed is interrupted by more than some amount of milliseconds, the car will stop and a human will be dispatched to correct the situation if it persists. This remote human has large amounts of data for them to understand the various known complexities of each road segment for each trip. Which areas are tricky (higher probability of needing to stop or have a poorer ride experience), which are newer (recent road changes), which are new to the segment (less traveled, no recent travel, less confidence)...etc
I've worked with Cruise on their AI infrastructure and I'm aware of their approach. It is largely documented
here. What is somewhat unknown to me as far as how it compares with Tesla is:
1. How much does it cost to outfit each vehicle from vanilla? I'm guessing about $100k in 3rd party hardware (sensors, compute, storage, communication, bespoke designing and labor) as I've done this work. Tesla's cost is >10x less and they make nearly all of them from scratch (vertically integrated at very low levels). Cruise will bring this cost down once they have a solution that can work without a city/region geofencing.
2. What is the current status of their AI to predict and have high confidence in corner cases for their geofenced areas? If anyone knows of recent videos to demonstrate this, I'd appreciate a link (I've looked on youtube and can't find anything recent within the past 6 months). With geofencing you can deterministically limit/cap/force your edge cases and fine tune how you deal with them. And the remote supervising human in-the-loop is simply not scalable. Geofencing is also done to ensure you don't take on unmitigated risk with a given number of possible road segments and complexities. For instance, your system has to ensure that the supervising human is attentive/aware and is able to see/hear as well or better *AS IF* they were actually behind the wheel at all times or the vehicle needs to safely stop.
Currently, with Tesla's approach, which is all road segments in given countries, I'm seeing issues with corner cases like complex left hand turns with or without traffic, but not much else. My hope is that unified vector space will be a huge leap forward to making this better. I'm still a huge proponent of basic, fundamental, structural 'near real-time' road segment data to augment real-time camera data during runtime prediction calcs. This can only be accomplished with unified vector space as well as turning camera data into ground truth, but there's an
Ashok patent for that!
3. How is GM's involvement going to impact the progress that Cruise is making? The less they are involved the better. Legacy ICE thinking will only hamper/dissuade/slowdown the progress of autonomous driving innovation...