Tesla Release notes | Laymens terms and how it applies to the experience |
-Upgraded the Object Detection network to photon count video streams and retrained all parameters with the latest autolabeled datasets (with a special emphasis on low visibility scenarios). Improved the architecture for better accuracy and latency, higher recall of far away vehicles, lower velocity error of crossing vehicles by 20%, and improved VRU precision by 20%. | Object detection is taking the output of the Occupancy map and identifying, classifying and probably prioritizing each item as an "Object of Interest" that is tracked by the Occpancy Flow map. This is the new way that 'raw-as-you-can-get-from-the-camera-sensor' photons are digitized to pixels and then turned into 3d representation of the world.
The autolabled dataset is what the NN is trained from which is constantly being refreshed as new data comes in from the fleet and trained on huge GPU clusters. Tesla has opted to prioritize "low visibility scenarios" like fog, mist, low light, blinding sun rays...etc. The new dataset requires retraining of the NNs and Tesla has done this on "all paramerters", which means everything that the NNs are looking for has been essentially reset and is new to the model. Nothing has been kept or held over from the last trained model and Tesla has opted to start completely fresh.
The new architecture refers to Tesla changing how the NN layers are built. NNs have many layers each with different properties for inputs and outputs. These layers could have been potentially modified to interact differently making them more efficient. Common practice in building out NNs.
The reason for all of this is to get better at detecting cars that are at the very limit of what the cameras can capture, get better at estimating speed of vehicles in crossing lanes and decreasing false positives (aka reduction of seeing things that aren't there; ghosts or improper slowdowns) for VRUs (pedestrians, bikes...Vulnerable Road Users) |
| |
-Converted the VRU Velocity network to a two-stage network, which reduced latency and improved crossing pedestrian velocity error by 6%. | Ego should not wait so long after a ped has crossed in front. And also should be more confident to proceed prior to ped crossing in front. |
| |
-Converted the NonVRU Attributes network to a two-stage network, which reduced latency, reduced incorrect lane assignment of crossing vehicles by 45%, and reduced incorrect parked predictions by 15%. | Ego should be able to make turns faster when cars are in other lanes than the lane that ego is entering or crossing. Ego should be able to tell when a car is parked vs waiting in traffic or waiting for it's turn to proceed |
| |
-Reformulated the autoregressive Vector Lanes grammar to improve precision of lanes by 9.2%, recall of lanes by 18.7%, and recall of forks by 51.1%. Includes a full network update where all components were re-trained with 3.8x the amount of data. | Huge update to the lane connectivity graph to make the car perform better when lanes intersect, merge and diverge. Ego should not miss as many turns due to confusing lanes. |
| |
-Added a new "road markings" module to the Vector Lanes neural network which improves lane topology error at intersections by 38.9% | Road markings are painted words and symbols like "Stop", "Turn Only", a turn line or turn and strait lines. Ego should miss less turns due to being in the wrong lane |
| |
-Upgraded the Occupancy Network to align with road surface instead of ego for improved detection stability and improved recall at hill crest. | Ego will understand the world much better as it comes over a hill and have less unnecessary slowdowns |
| |
-Reduced runtime of candidate trajectory generation by approximately 80% and improved smoothness by distilling an expensive trajectory optimization procedure into a lightweight planner neural network. | Ego will determine which way is the best way to go much faster and smoothness of turning is improved as it is now controlled by a fast, simple NN |
| |
-Improved decision making for short deadline lane changes around gore areas by richer modeling of the trade-off between going off-route vs trajectory required to drive through the gore region | Ego will go out of the lane in order to make lane changes when lanes diverge or merge when it needs to happen quickly. Gore region is simply the area between two lanes when they diverge or merge |
| |
-Reduced false slowdowns for pedestrians near crosswalks by using a better model for the kinematics of the pedestrian | Like #2, this improves how ego deals with peds crossing where it will 'do the right thing' more often. In other words, it will go when it should and NOT go when it shouldn't. |
| |
- Added control for more precise object geometry as detected by the general occupancy network. | Guessing here as this description is pretty vague. But, this seems like the first mention of using smaller voxels for parking situations. As it becomes more precise at close distances and slower speeds, it will be able to more accurately measure distance. Should become way better than the sonar pucks as this control gets better |
| |
- Improved control for vehicles cutting out of our desired path by better modeling of their turning/lateral maneuvers thus avoiding unnatural slowdowns | Oh thank goodness for this one. Ego should not wait for a car to completely leave the lane in order to proceed. Currently, ego will wait until the target car is completely out of the lane of travel when the target car makes a right turn. |
| |
-Improved longitudinal control while offsetting around static obstacles by searching over feasible vehicle motion profiles | Ego should proceed at a more natural speed when a car passes in a tight scenario. Currently, ego will come to a complete stop on unmarked roads. So hoping this is much improved. |
| |
-Improved longitudinal control smoothness for in-lane vehicles during high relative velocity scenarios by also considering relative acceleration in the trajectory optimization | This is another big one and really needed. Ego will now act more normally when coming up to a stopped car that has just started accelerating as well as when during the #11 scenario of a car turning right and is braking for the turn. |
| |
-Reduced best case object photon-to-control system latency by 26% through adaptive planner scheduling, restructuring of trajectory selection, and parallelized perception compute. This allows us to make quicker decisions and improves reaction time. | Just like it says in the last sentence...Ego will overall be making faster decisions. |
| SUMMARY |
| Nearly all of this stuff is for comfort, rather than safety. Currently, ego is acting slow and too conservative. These changes allow ego to act much more human-like by making quicker, more confident and smoother decisions. Lots of these updates are targeted at less unnatural slowdowns. |
| 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 |
| When peds are around the lane of travel |
| What ego does when on unmarked roads, should continue at a more natural speed when another car passes |
| What ego does when in a tight space while a car is coming the opposite way, should continue at a more normal speed |
| Does ego make faster decisions overall? |
| Does ego make smother turns? |
| Are there less unnatural slowdowns overall? |
| Does ego make turns into a lane even though there is a car in the adjacent lane? Hasn't done this well in the past |