V11 | |
Tesla Release notes | Laymens terms and how it applies to the experience |
- Enabled FSD Beta on highway. This unifies the vision and planning stack on and off-highway and replaces the legacy highway stack, which is over four years old. The legacy highway stack still relies on several single-camera and single-frame networks, and was setup to handle simple lane-specific maneuvers. FSD Beta's multi-camera video networks and next-gen planner, that allows for more complex agent interactions with less reliance on lanes, make way for adding more intelligent behaviors, smoother control and better decision making. | Single stack! FSD on highway! Occupancy map everywhere! Demonstrates huge confidence in this NN as the speeds are much higher. Occupancy flow of high speed kinematics is also very interesting. |
| |
Improved Occupancy Network's recall for close by obstacles and precision in severe weather conditions with 4x increase in transformer spatial resolution, 20% increase in image featurizer capacity, improved side camera calibration, and 260k more video training clips (real-world and simulation). | OccNet improved recall means less false negatives, meaning less times the car thinks it sees something that is NOT actually there.
Improved precision is less false positives, meaning less times the car thinks something IS NOT there, when in reality IS there.
The improvement in precision is for poor weather, meaning that FSD will operate in these conditions much better and it is therefore less likely to put the warning message about "reduced functionality due to poor weather detection". This is great for Seattle!
The OccNet 4x increase in transformer spatial resolution simply means that it can see 400% more volumetric detail, meaning that it should be much better at seeing smaller things that are close up and see things farther away. 20% increase in image featurizer capacity I think means they can support 20% more features in the OccNet, which means they can hopefully now recognize more road signs. |
| |
Improved merging behavior by leveraging lane geometry and lane bounds, association with coarse map information and better gap selection algorithms, allowing for smoother and safer experience. | I'm guessing that lane geometry (curve radius, width dynamics, intersecting locations) and lane bounds (precise location in the OccNet) were previously NOT used in lane merging, nor coarse map information (this was talked about at AI Day and includes things like 'Is_Bus_lane' and other deterministic markings/signs), but was already a NN as it says 'better gap selection algos'. Seems that we could then call it a Lane Merge Net (LMN). I'm guessing that they've been working on this a while as the same LMN would need to be used everywhere, so they needed to train it on high speed and very slow speeds. This sounds very difficult, so lets hope it is a good experience out of the gate! |
| |
Added highway behavior to offset away from blocked lanes and generic obstacles like road debris while also adding a smooth hand-off between in-lane offsetting and lane changing. | This seems simple, but as I read this it sounds like they have at least two NNs (in-lane offsetting - ILO NN and lane changing - LC NN). So, I read this as they have these input NNs into the neural planner. The neural planner then can seemingly plan better (with more confidence and NOT changing at the last second jerking the steering wheel). Will be interesting to experience what objects are avoided and how well it senses and adapts to blocked lanes. Hopefully construction zones and "Road Closed" signs trigger a route change and it is handled smoothly. |
| |
Improved speed-based lane change decisions to better avoid slowing down traffic in fast lanes, and interfere less with navigation. | Pretty straightforward, will be interesting to see how humanlike this feels. Hopefully less missed exits and entrances which I'm assuming they mean by saying 'interfere less with navigation'. |
| |
Reduced sensitivity for speed-based lane changes in CHILL mode. | Again, straightforward. Chilled out lane changes, not last minute, nice and smooth, get in the correct lane and stay in the lane. |
| |
Improved lane changes to allow higher jerk maneuvers if required to stay on-route or to move away from lane blockages. | This sounds like allowing FSD to operate outside the lane boundaries on occasion when merging like humans do when it is safe. |
| |
Improved smoothness at highway lane splits by being less strict about centering between lane lines and allowing lower jerk maneuvers, where safe to do so. | More humanlike highway interchange behavior like humans drive, hopefully less lane centering when going around a bowtie interchange. I think the lower gerk maneuvers could be in stop and go traffic, to smooth out the behavior. I find it nearly unusable in some stop and go traffic, so I hope it applies in those situations. |
| |
Reduced latency of trajectory optimization by 20% on average, without sacrificing behavior, by leveraging numerical tricks for more efficient computations. | Faster reactions to up-coming road and object dynamics. It doesn't say specifically, but this could apply to both the FSD car and cars around it. It is great to see this get faster as there are times when you just want FSD to react to something and then it seemingly reacts but a second too late. Let's hope it reacts more like a human! |
| |
| SUMMARY |
| This build is for single stack integration. Making FSD work at speeds higher than 55 MPH zones and hopefully acting more humanlike at higher speeds like anything over 45 MPH (usually the limit where there are crosswalks and predictable VRUs) |
| Things to test/'pay attention to' with this build |
| When cars make rights in front of ego, should be much more human-like and not cause someone behind to honk |
| When cars leave the lane of travel while braking, should be much more human-like and not cause someone behind to honk |
| All types of merging, lane changing should improve as well as blinker behavior |
| Try Chill mode and see if there is a noticeable difference from other modes |
| Try lots of highway stuff and see how it handles on and off ramps, bowtie interchanges and traffic. |
| Does ego make faster decisions overall? |
| Does ego make smother turns? |
| Are there less unnatural slowdowns overall? |