"Full Autonomy Feature Complete"???
The Tesla Autonomy Day last week Monday April 22 was an interesting affair. Two things are clear. The fact that Elon Musk has an inspiring vision is an understatement. And, that his predicted timelines for full self-driving (FSD) are credible is somewhat of a stretch.
Rather than the softball ones lobbied by the analysts, questions I would have liked to see answered:
Does each car have its own machine learning model and does it do model training as well? What is the need for a lot of computation power in the car?
Is each car learning in real-time and adapting itself from what its driver is doing?
There have now been at least two fatalities in strikingly identical situations: a white trailer-truck parked perpendicular to the freeway. Why did the machine learning model not learn from over a two-year old incident?
It’s totally possible that Tesla can pull off full autonomy. But one thing is clear. It’s not happening in the timeline being promised. We are a long ways away from that. Only once full autonomy has already been available with hands on steering wheel for several months is there any chance for regulatory approval for hands-free use. To talk about that today is jumping the gun. It may happen next year, or it may happen in 10 years. Elon Musk has no idea.
My team built a voice-driven enterprise virtual assistant for meetings and collaboration at Cisco Webex. Certainly nothing of the complexity of self-driving tech (nobody will get hurt if system fails), this was nevertheless a challenging undertaking with wake word models, speech recognition, natural language processing models, natural language generation, dialogue management, user interface refinements etc. After developing for a year, it was clear we were behind our aggressive timeline to release in 3 months. A release is within reach if the product has been in internal use for several months and people are generally happy with it. This stage can take a few months or years to get to. To talk about a release timeline prior is immature and reckless, specially when people’s lives are at stake.
Ever since Autosteer debuted in 2016, a few features have been added since, with the pace accelerating in 2019:
Auto Lane Change
Navigate on Auto Pilot: freeway ramps
Traffic light warning
With talk of Hardware 3, there does look like there is an increased feature momentum, but the road to FSD has to go through a lot of hurdles before it’s ready. At the minimum, these scenarios have to be handled:
City streets: Freeway roads are the easiest to navigate with AutoPilot due to best quality and more consistency. Internal roads add a lot more complexity and variation albeit at lower speeds. Red lights, stop signs, roundabouts, missing lane markings, left turn on yield, U-turns, dogs, cats, deer, bears, flooding, plastic bags, paper, falling objects etc.
Parking lots
Home Garage: Need accuracy to within a few a couple of inches for this to work properly without warning sirens going off every time. Need to handle steep ramps.
Construction zones: follow human instructions
Railway crossing
School zone: follow instructions of crossing guard
Ambulance / fire truck
Aggressive human drivers: This can get tricky. Humans can find a way to game the system so that a Tesla gets stuck at a stop sign or a left turn on yield, since it will always err on the conservative side.
Aggressive pedestrians: This is a whole other set of scenarios.
I expect some other features will be available before FSD is commonplace:
AutoSupercharging: Once in the Supercharger parking lot within 100 feet, driver gets off the car, pushes a button and goes away to pass 40 minutes. The car parks itself in a charging spot, charges and then parks in another spot waiting for driver.
AutoValet: Speeds are slow, but navigation can be super tricky with narrow multi-level ramps that go up and down several stories. Enables cars packed like sardines. Cars that park themselves and fetch themselves.
I use AutoPilot features so often now that it feels odd to drive without it. Full-serve driving is within reach. But each of the features above require a whole series of mini-steps before they will be practical. Several months for each in the fastest timeline I can imagine. Even if Tesla significantly ups its feature rollout, I don’t see full-self driving happening before 2024.