Tesla has just rolled out its ‘Full Self Driving (supervised)’ system to right-hand drive markets including Australia, and we recently had the chance to test the autonomous vehicle technology around Brisbane in a Model Y, which drove me over 50km without needing to touch any controls (well, almost).
Tesla will roll out FSD (supervised) to Australian customers with Hardware 4 (HW4) compatible vehicles from September onwards. The package will be available either as a 30-day free trial to new customers, or as a paid monthly subscription. Those properly convinced by the tech will also be able to fork out $10,100 to buy it upfront. See our separate news article for more details.
So how did it all go? Well, Tesla has been fine-tuning the training model behind FSD for some time in Australia and New Zealand. It needed to learn all the intricacies of Australian roads, including quirks like hook turns in Melbourne.
As part of this demonstration Tesla invited media to Queensland, where we were able to drive from Mount Cotton to the Brisbane CBD and back in FSD (supervised mode). While it isn’t a fully autonomous system (it’s only classed as a Level 2 system, which requires full driver control at all times), it doesn’t require periodic steering intervention like most other ADAS systems on the market.
It simply uses a camera within the cabin that monitors the driver to ensure they are paying attention. If the driver is distracted it will start prompting them to pay attention before disabling FSD (supervised) for the duration of the entire drive.
The curated Tesla route took us through some suburban streets around Mount Cotton before joining the freeway and heading into the CBD and back again. I didn’t want to be a pessimist, but I thought it was safe to assume Tesla tested this drive route countless times before inviting media.
So I decided to change the route once we got near the city. I diverted us to a Bunnings store to see how the Tesla would cope with a typical carpark full of dual-cab utes and people with wandering attention.
Before we get to the Bunnings experience, it’s worth pointing out how the rest of the drive went. It was a truly daunting experience because unlike every other semi-auto driving system available at the moment, you genuinely don’t have to touch any controls.
FSD (supervised) steers, brakes and accelerates for you, and will also change lanes and indicate for you.
Once you’ve spent five minutes or so behind the wheel the car starts feeling very natural and confident in its decisions. It doesn’t wander about in its lane like some lane-centring systems do and it’s very certain in the way it stays within its lane around other traffic.
It feels like a lot of attention has been paid to the way it interacts with traffic around it.
Tesla’s USP with this technology is that the US automaker only uses cameras to achieve ‘full self driving’. Most other brands will use a blended mixture of cameras, radar and LiDAR. Tesla says the key point of difference is that while having lots of sensors is great, they are useless if you can’t use software to interpret their data.
Tesla trains its model with what it calls ‘good drivers’. It’s a set of drivers used internationally to teach the car how it should merge, for example, or how much space is sufficient for changing lanes between other moving vehicles. They’re all decisions that it only wants to learn from people intentionally driving well.
It’s the main reason it can’t collect swathes of data from its customers. Because there are such a huge variety of driving styles and risk tolerances, it would make the system unsafe and incapable of making correct decisions.
But it’s not perfect. At one point on the freeway we slowed to 80km/h for a roadworks zone. Despite passing two 100km/h signs after the roadworks zone, the vehicle remained at 80km/h until we passed a third sign several kilometres further down the road.
It also failed to merge into the correct turning lane in the Brisbane CBD and, after passing the turn-off, it attempted to veer across traffic to go where it was meant to go. At this point I had to intervene and yank the wheel away from the system (you can see all of this in our video above).
It did impress when it came to a family of ducks crossing the road, though. On an 80km/h road it slowed down as we approached the 10 or so ducks crossing the road.
Now, what about Bunnings? On approach to the store there was a leaf blower and pedestrian that got in the way of the car. It successfully recognised both of them and slowed down until they moved out of the way. It then navigated its way through a fairly tight carpark. It was all going great until it sailed through a give-way sign and failed to yield to another vehicle that had right of way.
It’s these types of incidents that don’t really instill confidence in the system. If I was sitting as a passenger in the car and it was doing all the driving on its own, I’d feel constantly stressed about how it was handling situations.
So what are truly autonomous cars like in comparison? I had the chance to ride in several Waymos in Los Angeles recently and, while I was a bit reluctant at first, the experience was pretty cool. Waymos, unlike Teslas, use a litany of LiDAR, radar and camera vision to sense what’s around them.
There’s nobody in the driver’s seat at all, so it’s not uncommon to see them driving around empty as they travel to their next fare. The Waymos I took felt very confident on the road and drove exactly like a human would. They also weren’t overly influenced by some of the crazy LA drivers and didn’t flinch when cars would veer into their lane ahead.
So how would I rate the overall Tesla FSD (supervised) experience? I think it’s a great starting point for supervised ‘self-driving’. But I wouldn’t feel confident being in the passenger seat or in the back seat while it was driving. There were too many of those situations that could have ended up being close calls if I wasn’t there to take control.
This may be a limitation of solely relying on camera vision or it could be a training model limitation. Both of these things can be improved over time with smarter software, so there’s scope for gradual improvement. I wouldn’t be against the introduction of more sensors – in my view with the correct software stack and data interpretation, more sensors is always better than fewer of them.
Either way it was a fun experience.
Click the images for the full gallery