– Upgraded the object detection network to photon count video streams and retrained all parameters with the latest auto-labeled dataset (with particular focus on low visibility scenarios). Improved architecture for better accuracy and latency, higher recall for distant vehicles, 20% lower speed error for crossing vehicles and 20% improvement in VRU precision.
– Converted the VRU Velocity network to a two-level network to reduce latency and improve crosswalk speed error by 6%.
– Converted the Non VRU Attributes network to a two-level network to reduce latency, reduce incorrect lane assignments for crossing vehicles by 45%, and incorrect parking predictions by 15%.
– Restructured autoregressive vector lane grammars to improve lane precision by 9.2%, lane recall by 18.7%, and branch recall by 51.1%. Includes full network updates with all components retrained with 3.8x the amount of data.
– Improved lane topology errors at intersections by 38.9% by adding a new “road markings” module to the Vector Lanes neural network.
– Upgraded Occupancy Network to align to the road instead of ego for better detection stability and better recall on hilltops.
– Reduced runtime of candidate trajectories generation by about 80% and improved smoothness by extracting the costly trajectory optimization procedure into a lightweight planner neural network.
– Improved decision making for short deadline lane changes around gores with rich modeling of the trade-off between going off route and the trajectory required to get through the gore area
– Reduce false deceleration for pedestrians near crosswalks using a better model for pedestrian kinematics
– Added control for more accurate object geometry detected in common occupied networks.
– Improves control for vehicles going off course by better modeling turn/lateral maneuvers to prevent unnatural deceleration.
– Improved longitudinal control while canceling around static obstacles by searching for possible vehicle motion profiles
– Improved longitudinal control smoothness for in-lane vehicles in high relative speed scenarios by taking relative acceleration into account in trajectory optimization.
– Reduced best-object photon-controlled system latency by 26% through adaptive planner scheduling, restructuring of trajectory selection, and cognitive computing parallelization. This allows us to make faster decisions and improve reaction times.
– Introduced native support for model parallel neural network inference by sharing intermediate tensors across SOCs to improve road edge and road line prediction consistency through changes in the TRIP compiler, inference runtime, and interprocessor communication layer.
– Improved traffic control behavior in dense intersection areas by improving connection logic between traffic lights and intersections.
To share your feedback, hit the “Record Video” button in the top bar UI. At the touch of a button, the car’s exterior camera shares a short VIN-related Autopilot snapshot with the Tesla engineering team to help improve FSD. I can’t see the clip.