In an try to be the primary to deploy a really fully autonomous driving system later this 12 months, Tesla is collecting corner cases that its vehicles encounter on the road with its Autopilot system.
We now get an interesting example with a visualization of how Autopilot interprets and reacts to a goose crossing the road.
At a presentation about Tesla’s latest self-driving and computer vision capabilities last month, Tesla Sr. Director of AI Andrej Karpathy said that they’re currently working to discover and react to corner case scenarios that vehicles encounter on the roads.
Karpathy says that they’re pooling numerous different rare events from their entire fleet of vehicles equipped with cameras so as to teach their Autopilot system to react to them.
Earlier this month, a Tesla owner claimed that Autopilot stopped for a rabbit on the road caught on a dashcam video.
On the time, we noted that the accuracy of the owner’s claim is unclear and that a visualization of what the Autopilot can see, like hacker verygreen has been producing, could be useful in this case.
While green couldn’t produce a rabbit out of his hat, a goose did appear on the road while the Autopilot data was being collected and he produced the next visualization using that data:
Well, this ain’t no rabbit, but we will clearly see Autopilot doesn’t respect Canadian Geese, it just doesn’t respect them. pic.twitter.com/dlBOIFmnaw
— green (@greentheonly) May 21, 2019
As you’ll be able to see, Tesla’s Autopilot doesn’t detect the goose for probably the most a part of the clip.
It does eventually detect and classify it as a pedestrian for a fraction of a second.
In the course of the second a part of the video, the goose isn’t classified, but Autopilot seems to know that it will possibly’t drive over it:
The green space is where Autopilot detects as secure to drive while the clear box, which appears over the goose, is interpreted as not appropriate to drive over.
Electrek’s Take
There’s numerous talks at once about whether or not Tesla’s goal of achieving full self-driving by the tip of the 12 months is realistic.
I feel the consensus is that it’s not and I are inclined to agree, but I also think that Tesla is making numerous progress on the AI and machine learning front.
If that progress proves to be exponential, we could have a surprise on the complete self-driving front.
Nevertheless, those corner cases, like animals crossing the road, looks like a particularly hard problem to unravel.
While it looks just like the Tesla vehicle would have run over the goose on this video if it was on Autopilot, I see some potential based on the visualization.
Within the meantime, drivers must ensure that they’re being attentive and able to take control in any respect times.