Tesla has just rolled out the latest Full Self-Driving Beta software update version 10.6 (firmware ver: 2021.36.8.9). According to the release notes of the 10.6 update, this latest installment has made significant improvements in the architecture that detects other vehicles on the road.
Tesla also claims an improvement in visibility network precision, thus minimizing the mean relative error by 18.5%. Considering the size of Tesla’s neural network, eliminating the error rate this much is quite phenomenal for a sub-version update.
In addition to the above-mentioned improvement in moving objects, Tesla has also made a 17% improvement in static object (e.g. cones, dustbins, boxes, poles) detection during a turn on high curvature and when the environment is relatively dark at night time.
Let’s explore the detailed improvements Tesla has achieved in the FSD Beta 10.6 update.
FSD Beta 10.6 Release Notes (2021.36.8.9)
- Improved object detection network architecture for non-VRUs (e.g. cars, trucks, buses). 7% higher recall, 16% lower depth error, and 21% lower velocity errror for crossing vehicles.
- New visibility network with 18.5% less mean relative error.
- New general static object network with 17% precision improvements in high curvature and nighttime cases.
- Improved stopping position at unprotected left turns while yielding to oncoming objects, using object predictions beyond the crossing point.
- Allow more room for longitudinal alignment during merges by incorporating modelling of merge region end.
- Improved comfort when offsetting for objects that are cutting out of your lane.
Since this is a relatively fresh update and most of the Tesla Early Access Program members are still downloading or waiting for it to reach their cars, video tests are scarce.
However, YouTuber James Locke is one of the earliest Tesla owners to get this update and he has posted a short test video after getting the update.
James drove his Tesla Model 3 on FSD Beta 10.6 at night time as soon as he got the update. Since Tesla says it has improved nighttime static object detection, driving must have gotten better in this area. However, It is not clear if the cars parked on the side of the road also count as static objects when the computer renders them as vehicles.
In a few instances in James’ video, his Model 3 on FSD Beta 10.6 crept away too more on some intersections, which looked like a lack of confidence. I also noticed that the car took an unprotected left turn a bit aggressively while a vehicle was still coming from the left, although the oncoming vehicle was at a distance, this did not look like a very good move.
Most of this test drive was not very challenging as traffic on the roads was low due to late hours. We will be covering more video tests as they become available with the FSD Beta 10.6 update.