Enabled faster identification of red light runners by evaluating Reduced latency when starting from a stop by accounting for lead Upcoming map speed changes, which increases the confidence of Improved speed when entering highway by better handling of Lane, especially in intersections or cut-in scenarios. This helps with objects turning into or away from ego's
Rate by incorporating yaw rate and lateral motion into the likelihoodĮstimation. Improved object future path prediction in scenarios with high yaw Improved recall of object detection, eliminating 26% of missingĭetections for far away crossing vehicles by tuning the lossįunction used during training and improving label quality. Trajectory estimation used as input to the neural network. Improved velocity error for pedestrians and bicyclists by 17%,Įspecially when ego is making a turn, by improving the onboard Tokens participate in the attention operations of the autoregressiveĭecoder and by increasing the loss applied to fork tokens during Increased recall of forking lanes by 36% by having topological Optimization to focus more on areas where finer control is essential. Improved accuracy of stopping position in critical scenarios withĬrossing objects, by allowing dynamic resolution in trajectory Might cross ego's path, regardless of presence of traffic controls. Enabled creeping for visibility at any intersection where objects Improved recall of animals by 34% by doubling the size of the To allow for smoother stops when protecting for potentially Made speed profile more comfortable when creeping for visibility,
Information bottlenecks in the network architecture wereĮliminated by increasing the size of the per-camera featureĮxtractors, video modules, internals of the autoregressive decoder,Īnd by adding a hard attention mechanism which greatly improved
#Sentry self storage full
Improved geometry error of ego-relevant lanes by 34% andĬrossing lanes by 21% with a full Vector Lanes neural network Improved understanding of pedestrian and bicyclist intent based on Reduced false slowdowns near crosswalks. Objects present and also improves yielding position when they are This reduces false slowdowns when there are no relevant
Increased smoothness for protected right turns by improving theĪssociation of traffic lights with slip lanes vs yield signs with slip 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
0 kommentar(er)
