Velocity estimates for far away crossing vehicles by 20%, while velocity, acceleration, yaw rate) where networkĬompute is allocated O(objects) instead of O(space). Upgraded to a new two-stage architecture to produce object Outlier rejection, hard example mining, and increasing the dataset Also, improved ground truth with semantics-driven This temporal context allows the network toīe robust to temporary occlusions and enables prediction of Upgraded Occupancy Network to use video instead of imagesįrom single time step. This required predicting velocity at every 3D Precise object shapes that cannot be easily represented by aĬuboid primitive. Added control for arbitrary low-speed moving volumes from That are entering or waiting inside the median crossover region with Finally, improved interaction with objects Also improved lateral profileĪpproaching such safety regions to allow for better pose that aligns Jerk, to mimic the harsh pedal press by a human, when required to This was done by allowing optimisable initial The presence of high speed cross traffic ('Chuck Cook style' Profile when approaching and exiting median crossover regions, in Improved unprotected left turns with more appropriate speed Smoothness while also allowing a more accurate response during This allows better downstream controller tracking and In a trajectory that is a more accurate model of how the vehicle Well as acceleration and brake commands to actuation. Trajectory planner now independently accountsįor latency from steering commands to actual steering actuation, as Through better modeling of system and actuation latency in Improved overall driving smoothness, without sacrificing latency, General way that adapts for road changes. Good as someone driving their own commute, yet in a sufficiently

This provides a way to make every Autopilot drive as

Lane topology compared to the previous model, enabling smootherĬontrol before lanes and their connectivities becomes visuallyĪpparent. This architecture achieves a 44% lower error rate on Neural network which fuses features extracted from the video Added a new 'deep lane guidance' module to the Vector Lanes

Autotap software keys