Abstract: In some embodiments, a method for autonomous navigation of an unmanned aerial vehicle (UAV) is provided. The UAV determines a tracked position using at least one positioning sensor of the UAV. The UAV captures an image using a camera of the UAV. The UAV determines a visual position confidence area using the captured image. The UAV checks the tracked position using the visual position confidence area to determine whether the tracked position is accurate. In response to determining that the tracked position is not accurate, the UAV causes a corrective action based on the visual position confidence area to be taken.

Abstract: A method for perception validation of an aerial vehicle includes: acquiring an image of a ground area with an onboard camera system of the aerial vehicle, generating an above ground altitude (AGL) estimate with a neural network trained to output the AGL estimate in response to the image fed as an input to the neural network, generating a motion estimate or a position estimate based upon sensor data output from a sensor disposed onboard the aerial vehicle, and cross-validating the motion or position estimate against the AGL estimate.

Abstract: A technique for processing delivery aborts by a UAV delivery service includes: acquiring a delivery zone (DZ) image of a delivery destination including a DZ for a package being delivered to the delivery destination by a UAV; determining to abort a delivery mission for the package based upon the DZ image; converting the DZ image to a vector embedding; performing a similarity search on a vector database using the vector embedding, the vector database storing reference images of other delivery destinations indexed to reference vector embeddings and outcome attributes describing delivery outcomes associated with the reference images, and wherein the similarity search identifies a subset of the reference images deemed to have a threshold similarity to the DZ image; and prompting a vision language model with the DZ image and the subset of the reference images to provide an abort explanation or to determine a delivery disposition for the DZ.

Abstract: A technique for provisioning UAVs to operate in a specific geographical region includes acquiring reference aerial images of the specific geographical region from a secondary source. The reference aerial images have at least one style characteristic that is distinct from that of operational aerial images acquired by the UAVs during operation. The reference aerial images are transformed into training images that adopt a style of the operational aerial images using a neural style transfer. A neural network that facilitates vision-based navigation of the UAVs is trained using the training images. The neural network is trained to operate in the specific geographical region prior to operation in the specific geographical region.

Abstract: A technique for informing navigation of a UAV includes storing an asset map of a ground area, wherein the asset map includes a reference aerial image of the ground area annotated with labels describing reference objects depicted in the reference aerial image; acquiring a current aerial image of the ground area with an onboard camera system of the UAV while the UAV is flying above the ground area; mapping correspondences between the reference aerial image and the current aerial image using a homography estimating tool executing onboard the UAV; analyzing the current aerial image with an object detection model to detect a first object positioned at the ground area; and validating or informing a detection of the first object by the object detection model based on the mapping of the correspondences.

Abstract: A method includes receiving an input specifying a starting location and a destination location for an aerial vehicle. The method additionally includes determining, based on the starting location and the destination location, an aerial path for the aerial vehicle to follow from the starting location to the destination location. The method also includes determining, based on the aerial path, a property of aerial image data, where the aerial image data is obtainable using the aerial vehicle while traversing the aerial path, and where the aerial image data represents an environment along the aerial path. The method further includes determining, based on the property, a path score associated with the aerial path, and outputting the aerial path based on the path score.

Abstract: In some embodiments, a computer-implemented method of communicating sensor data between an autonomous vehicle computing system and a remote computing system is provided. The autonomous vehicle computing system gathers sensor data generated by at least one sensor of the autonomous vehicle computing system. The autonomous vehicle computing system determines an amount of available bandwidth between the autonomous vehicle computing system and the remote computing system. The autonomous vehicle computing system retrieves a local portion of a split machine learning model corresponding to the amount of available bandwidth from a model data store. The autonomous vehicle computing system processes the sensor data using the local portion to produce an intermediate result. The autonomous vehicle computing system transmits the intermediate result to the remote computing system for processing of the intermediate result using a remote portion of the split machine learning model.

Abstract: A method includes receiving an input specifying a starting location and a destination location for an aerial vehicle. The method additionally includes determining, based on the starting location and the destination location, an aerial path for the aerial vehicle to follow from the starting location to the destination location. The method also includes determining, based on the aerial path, a property of aerial image data, where the aerial image data is obtainable using the aerial vehicle while traversing the aerial path, and where the aerial image data represents an environment along the aerial path. The method further includes determining, based on the property, a path score associated with the aerial path, and outputting the aerial path based on the path score.

Abstract: A computer-implemented method comprises receiving an image captured by a camera on an unmanned aerial vehicle (UAV). The image depicts an environment below the UAV. A feature mask associated with the image is generated via a machine learning model that is trained to identify and semantically label pixels representing the environment depicted in the image. One or more reference tiles associated with the environment are retrieved. The reference tiles are associated with particular geographic locations and specify semantically labeled pixels representing the geographic locations. The semantically labeled pixels of the feature mask are correlated with the semantically labeled pixels of at least one of the one or more reference tiles to determine the geographic location of the UAV in the environment.

Abstract: A technique for managing an unplanned contingency landing of a unmanned aerial vehicle (UAV) includes determining by the UAV that the unplanned contingency landing is imminent, capturing aerial images of a ground area below the UAV with an onboard camera system as the UAV descends towards the ground area, semantically analyzing the aerial images to classify objects at the ground area into object classifications, depth analyzing the aerial images to determine above ground level (AGL) heights associated with each of the objects at the ground area, selecting a preferred landing site for the unplanned contingency landing that is coincident with one of the objects at the ground area based upon a contingency landing policy that ranks contingency landing sites based upon a combination of the object classifications and the AGL heights; and nudging the UAV towards the preferred landing site as the UAV descends towards the ground.

Abstract: Introduced here are diabetes management platforms able to examine glucose measurements associated with a subject to discover patterns indicative of similar behaviors/circumstances. More specifically, a diabetes management platform can acquire multiple data series generated by a glucose monitoring device that monitors the blood glucose level of an individual over multiple time intervals. The diabetes management platform can then apply a similarity algorithm to produce a distance measure (also referred to as a “similarity measure”) for each data series. For example, the diabetes management platform may, for each data series, perform dynamic time warping to produce a distance measure with respect to each of the other data series. The diabetes management platform can identify patterns indicative of similar behaviors/circumstances based on these distance measures.

Abstract: A method for perception validation of an unmanned aerial vehicle (UAV) includes: acquiring an aerial image of a ground area with an onboard camera system of the UAV, generating a semantic above ground altitude (AGL) estimate with a neural network trained to output the semantic AGL estimate in response to the aerial image fed as an input to the neural network, generating a motion estimate or a position estimate based upon perception sensor data output from a perception sensor disposed onboard the UAV, and cross-validating the motion or position estimate against the semantic AGL estimate.

Abstract: A technique for localization of an unmanned aerial vehicle (UAV) includes: acquiring aerial images of a terrain below the UAV with an onboard camera system of the UAV while the UAV is flying a mission along a preplanned route over the terrain; generating a current optical flow map based upon image pixel motion between consecutive images in a sequence of the aerial images; comparing the current optical flow map to reference optical flow maps stored onboard the UAV, wherein the reference optical flow maps are precomputed from a model of the terrain along the preplanned route; and determining a position in at least two lateral dimensions based on the comparing.

Abstract: A method for unmanned aerial vehicle (UAV) mission planning includes acquiring a target aerial image of a geographic area representative of the geographic area illuminated by one or more artificial light sources, identifying a location of the one or more artificial light sources based on the target aerial image, rendering a simulated aerial image representative of the geographic area illuminated by the one or more artificial light sources at night using a digital surface model of the geographic area, the location of the one or more artificial light sources, and an irradiance parameter for the one or more artificial light sources, identifying one or more regions within the geographic area as having sufficient lighting for UAV operation at night based on the simulated aerial image, and generating a mission plan for the UAV based on the one or more regions within the geographic area.

Abstract: A technique for processing delivery aborts by a UAV delivery service includes: acquiring a delivery zone (DZ) image of a delivery destination including a DZ for a package being delivered to the delivery destination by a UAV; determining to abort a delivery mission for the package based upon the DZ image; converting the DZ image to a vector embedding; performing a similarity search on a vector database using the vector embedding, the vector database storing reference images of other delivery destinations indexed to reference vector embeddings and outcome attributes describing delivery outcomes associated with the reference images, and wherein the similarity search identifies a subset of the reference images deemed to have a threshold similarity to the DZ image; and prompting a vision language model with the DZ image and the subset of the reference images to provide an abort explanation or to determine a delivery disposition for the DZ.

Abstract: A technique performed by UAV delivery system includes: arriving by a UAV of over a destination area; capturing a plurality of aerial images of a scene at the destination area with an onboard camera system of the UAV while flying above the destination area, wherein the aerial images capture the scene from a plurality of UAV vantage points offset from each other; optimizing weights of a generative neural network (GNN) using at least some of the aerial images as a training dataset to encode a volumetric representation of the scene into the GNN, wherein the weights are optimized by an onboard processing system of the UAV; and communicating the GNN with the weights optimized to a backend datacenter in communication with the UAV to transmit the volumetric representation of the scene over which the UAV flew without transmitting the aerial images themselves to the backend datacenter.

Abstract: A technique for avoiding a dynamic obstacle in a vicinity of UAV while the UAV flies a mission along a planned route includes: receiving a wireless radio signal containing information related to the dynamic obstacle in an audible speech format; translating the audible speech format to a text format; analyzing the information to determine an identity of the dynamic obstacle and determine whether the information relates to a flightpath, speed, or heading of the dynamic obstacle; in response to determining that the information relates to the flightpath, speed, or heading of the dynamic obstacle, determining an updated flightpath, speed or heading for the dynamic obstacle; comparing the updated flightpath, speed, or heading to the planned route; and in response to detecting a conflict between the dynamic obstacle and the UAV based on the comparing, performing an action by the UAV to avoid the conflict with the dynamic obstacle.

Abstract: A method includes navigating, by a UAV, to a delivery location in an environment; capturing, by at least one sensor on the UAV, sensor data representative of the delivery location; determining, based on the sensor data, a segmented point cloud of the delivery location, wherein the segmented point cloud defines a plurality of point cloud areas with corresponding semantic classifications; determining, based on the segmented point cloud, that a pre-selected delivery point at the delivery location satisfies a condition indicating that a descent path through a cylinder, the cylinder being centered above the pre-selected delivery point and having a radius of a particular lateral distance, does not intersect with any point cloud areas having semantic classifications indicative of an obstacle at the delivery location; and based on determining that the pre-selected delivery point satisfies the condition, initiating, by the UAV, a payload delivery operation towards the pre-selected delivery point.

Abstract: A method of collaborative analysis of a ground area by UAV delivery service includes acquiring first and second aerial images of the ground area. The first and second aerial images include depictions of objects at the ground area. A query including an encoding of the first aerial image is transmitted to a cloud-based neural network trained to identify objects. A motion of the UAV is tracked between acquiring the first and second aerial images. A response is received from the cloud-based neural network identifying one or more of the objects depicted in the first aerial image. An onboard neural network disposed on board the UAV is used to identify the objects at the ground area. The onboard neural network receives the response, an indication of the motion tracked between the first and second aerial images, and the second aerial image as input when identifying the objects.

Abstract: A method includes receiving an input specifying a starting location and a destination location for an aerial vehicle. The method additionally includes determining, based on the starting location and the destination location, an aerial path for the aerial vehicle to follow from the starting location to the destination location. The method also includes determining, based on the aerial path, a property of aerial image data, where the aerial image data is obtainable using the aerial vehicle while traversing the aerial path, and where the aerial image data represents an environment along the aerial path. The method further includes determining, based on the property, a path score associated with the aerial path, and outputting the aerial path based on the path score.