Back in April this year, Elon Musk simultaneously made the world excited and nervous by proclaiming their Full Self-Driving technology would be feature complete by the end of 2019.
During the presentation, Musk explained the reality of delivering FSD is actually going to take three steps to implement. There’s certainly been some progress this through OTA software updates, like Navigate on Autopilot that will automatically overtake slow cars (on the freeway).
Despite the progress, the reality is that we’re just 3 weeks out from the first deadline and the to-do-list seems mighty in that timeframe.
What’s left on the to-do-list
Musk has spoken about (and tweeted) about the fact they see autonomous driving being achieved by tackling three elements of travel. The highway stuff is largely done and they have millions of miles (or km) to prove that works. The second challenge is the low-speed navigation and you could argue an improvement to Smart Summon could tick that box.
The remaining and likely the most difficult is the building the intelligence to drive between those two environments.
Imagine the average daily commute. You pull out of your garage, navigate tight city streets, typically filled with lots of obstacles, signs, lights, speed zones, potentially roadworks, then route through to a highway onramp, exit the freeway, then navigate some more suburban streets to arrive at your office.
Some of this can be achieved today, but much of it can not and that’s what makes me nervous about hitting the timeline.
Traffic lights
Back in April, Musk spoke about traffic light support on Lex Fridman’s podcast. At the time Musk said his Tesla was already running a development version that stopped at traffic lights. Now, 8 months later and we still don’t have that.
In the past few days, Tesla has begun rolling out update 2019.40.2 that can read traffic lights and alert drivers if they are approaching a red light. Musk said on the podcast, this warning level was the precursor to the car stopping on its own. This means we’re probably close on this one, but it’s a tight timeline.
As with many of Tesla’s announcements, they aren’t explicit about the international availability of some of this, so while we’ve seen screenshots confirming the traffic light warnings, we don’t yet know if this feature will be available outside the US.
With many differences in traffic lights and intersections between countries, it is understandable that training the AI to cope with these variations would take longer, but when Tesla speaks generally about feature availability, that sets global expectations.
Stop Signs, Give Way signs
At the Tesla Autonomy Day, back on April 22nd, Tesla released a video of a Model 3, driving itself through city streets and stopping at Stop signs, waiting, then proceeding when clear, just as a human would. While that was amazing to see, again, we’re 8 months on from then and it seems that capability is still not ready for release.
Roundabouts
Every day, I travel through at least 4 roundabouts and this means tapping the brake to release Autopilot and manually navigating the dual-lane roundabouts. If I was to let the Tesla handle it, it’d happily drive me right over the center island, or into the car next to me.
For me personally, this would be the single biggest feature upgrade to the driving experience today and a critical part of achieving FSD. I get it, roundabouts are complex, you can have anywhere from 1 to 5 or more exists and humans are generally terrible at navigating them. Almost daily I see drivers cut lanes without indicating, incorrectly judge the space required to enter or just plainly fail to understand who’s meant to go first.
Driving on roads without good lane lines
My whole estate is fairly new and none of it has lane markings on the road. What it does have is grey, cement gutters that contrast the black tarmac well, so I’m surprised Autopilot can’t identify that as a lane.
Driving the Model 3 for around 4,000kms, I’ve noticed this a lot around town. While most roads are marked, those that aren’t are common enough that it really needs to be solved to make FSD viable in getting you from home to work and back.
Parking
Smart Summon is an impressive technology, but it currently only works on flat parking lots. The reality is that many of us park in multi-story carparks. These are complex and difficult, but delivering FSD really means meeting and exceeding human capability.
Every day humans navigate multi-level carparks and coming into Christmas, they’ll be insanely busy with plenty of obstacles to avoid. I don’t doubt Tesla can craft AI to cope with these scenarios, but the timeframe is very much in doubt when it comes to dealing with this challenge.
While we’re on the topic of parking, its generally accepted that Tesla’s auto-park (manually initiated) is average at best. There’s not many areas that Tesla lags behind other automakers, but this is certainly one of them.
Feature Complete (due end of 2019)
I really hope Tesla doesn’t use Navigate on Autopilot, Smart Summon and some stoplight recognition as the justification to claim they’ve achieved FSD feature complete.
What I hope is that Tesla’s ramp in production this year has enabled them to dramatically accelerate the data provided to their AI to learn from. As the engineers label more data and feed that to the AI models for each of the challenges listed about, it’s possible that we have an acceleration in the delivery of new FSD features.
Given the software rolling out right now focuses on something as trivial as automatic windscreen wipers, it’s difficult to have confidence Tesla will tick all the boxes listed above.
Driver does not need to pay attention (Q2 2020)
Assuming Tesla does pull a rabbit out of the hat and deliver the three pillars of FSD, the job is certainly not complete. As we’ve experienced with Autopilot, it does a great job, most of the time, but before we can relieve humans of their responsibilities completely, you have to be incredibly confident FSD can handle almost every edge case.
Musk detailed the ‘driver does not need to pay attention’ timeline as somewhere in the second quarter of 2020. That’s also an aggressive timeline to set, particularly given there’s no telling the volume of weird and wacky issues that may pop up across the globe.
Once this level is achieved, the nag to put pressure on the steering wheel can finally go away. Drivers could be tired and get home safely. Technically, the car could drive itself, so we could even let the car be the designated driver, but while the possibility to take control still exists (car has a wheel and pedals), regulators will still require the person in the driver’s seat, to be at a BAC of 0.5.
Regulator approval for FSD (limited locations end of 2020)
The third and final step of FSD is getting approval to do it. It is important to remember the scale of the challenge for Tesla here. Tesla is pioneering autonomous driving, others are working on it, but there’s no doubt in my mind, Tesla will be the first to reach Level 4/5 Autonomy where the driver is not required.
There is no United Nations for autonomy that you can go to and present your technology and safety data, Tesla will need to go to regulatory bodies for each state in each country they sell vehicles in. This is not only a huge logistical challenge, but there’s bound to be many uneducated that resist it and are scared by an idea this radical.
Musk said.. “the fleet wakes up with an over the air update” and by that he meant FSD will arrive in almost a million vehicles that have their HW3 platform.
Don’t be surprised if we see more protests from the likes of the Taxi lobby who’ll inevitably complain about silicon valley destroying more jobs. If you think they were annoyed by Uber stealing work from them, just wait till they’re not competing with other humans, but competing with driverless vehicles, it’ll be a disruption like we’ve never seen before.
Tesla aren’t yet making vehicles for themselves, but clearly have a plan to do so. Once the production ramp for Model 3, Model Y reaches a peak rate and demand from regular consumers flattens, Tesla will likely to continue making vehicles for themselves (at cost) and leveraging their FSD capabilities to go out and make more money for the company as robotaxis.
In addition there is nothing on the horizon to auto drive when rain, snow, or fog result in almost zero visibility. This can occur at any time and will immediately require manual takeover by a human occupant.