Tesla is currently madly building their Fully Self Driving AutoPilot system. Trying to go from self driving on freeways only, to self driving on city streets and in car parks, to full L4 self driving everywhere.
As a new Tesla owner, something occurred to me while driving around the other day. Why not add a new overlay to the main Tesla map that tells owners where Autopilot has the most trouble driving?
Currently AutoPilot is meant for highways only, but it does do quite well on city streets, at least the ones that are clearly marked and don’t deviate much from what a highway might look like.
A new overlay on the main Tesla map could not only help Tesla develop their AutoPilot system faster, but reward owners for helping to gather the data and make AutoPilot handle their local suburbs better in the process.
A Map Of Disengagement
Currently when you disengage AutoPilot and take over, Tesla logs the incident and (I’m guessing) uses that information to improve the system as a whole. After all, if you’ve had to take control then obviously the system is having troubles in that spot and something needs to be looked at or fixed.
In this way, Tesla already has a “map of disengagements” on record. A map that shows them which localised areas the system has trouble with and which ones are all good. What I’m proposing is to make this data visible for all owners to see and use.
This could take the form of a second Traffic map overlay. In some areas where the roads are well marked and AutoPilot does fine with no notable disengagements by users it could be marked as say, blue.
In other areas though, where the road isn’t so clear or where there’s strange corner cases like graffiti over signs, damaged roads, rubbed out lines etc you would see lots of disengagements by AutoPilot users and these areas would be highlighted as red or even black if particularly bad.
These would be the areas that people have perhaps tried AutoPilot and had to disengage it, taking over as the system obviously can’t handle it or just doesn’t do it properly for whatever reason.
A Teachable Experience
Once this overlay was live you could toggle it just as you do the traffic layer and see all the streets and areas around you that are red or black. Once owners see that the street down the road is “black” they will be encouraged to drive that path again.
As more and more owners manually drive these red or black “bad” areas, Tesla would explicitly gain recorded data on how the system should be driving there. It wouldn’t be AutoPilot driving these “bad” areas, it would be the owners teaching the system exactly how it should be handling these edge cases.
As more and more owners drive the same area, Tesla would gather more data of how different humans handle the route.
It would also get data on driving the same route at night, during the day, through wind, rain or sunshine. It’s in this way that Tesla could teach its neural net to drive a previously undrivable route basically without having to do anything.
The AI would automatically identify badly driven areas (through disengagements) then dynamically receive heaps of human driven examples or “correct solutions” on how to solve that badly driven area.
In this way, Tesla owners could further help to improve the AutoPilot system, especially for bad areas around where they live. Tesla also obviously benefits as it gets to automatically build up their FSD system with super important edge cases.
Gamification & Rewards
While Tesla owners are notorious for going the extra mile for Tesla, they could take this whole idea a step further and gamify it. For every kilometre you drive in red or black areas you could get points.
There could be leader boards, top 10 lists, local competitions and more. Prizes could be dolled out just like Tesla does for their referral program. Driving “red” zones might get you 1 point per kilometre driven while driving and gathering data in the the really bad “black” areas could get your 2 points per kilometre.
These points could be traded in for free Supercharger kilometres or other prizes and rewards. There might even be further options to have owners flag the locations with problems like “lane markings degraded” or “stop sign knocked down”.
This data could then also be forwarded onto the local councils so they could be informed about areas that need repair. The options are really quite endless.
Turning Data Sharing On
Obviously to take part in this feature you’d have to agree to let Tesla collect the recordings of you manually driving through these “bad” AutoPilot areas.
This in itself would mean more owners would be exposed to enabling the data sharing option which is turned off by default (as it should be) on Tesla’s. This option is also buried quite deep in the menu system and if you’re not a tech nutter going through every part of your new Tesla or reading all the Tesla guides/articles you could easily miss it.
As such, simply having this new “game” would call more attention to this data collection toggle and more people would be likely to turn it on further helping Tesla gather data.
This model of crowd sourcing local data is also something that has already proven to be really good. Google Maps and their Local Guides system is a similar model where they reward Google uses with prizes or badges for reviewing a certain number of places. Some have gotten 1 TB of free Google Drive storage for a few years in return.
Either way, I think the case for Tesla building a type of AutoPilot Game is quite clear. One of Tesla’s primary mission of building its Full Self Driving AutoPilot system would get a huge boost of data. This data would be for the most important edge cases and it would be of real life humans driving these routes showing the AI how it’s done.
Beyond that it’d increase the number of owners contributing to Data Sharing in general and make for a fun new game that all Tesla owners can participate in. A huge win all around!
Source: Alex Shoolman. Reproduced with permission.