Abstract: Systems and methods for payload management systems are illustrated. One embodiment includes a delivery mechanism. A mechanical linkage in the delivery mechanism includes rocking bars mounted to: a chassis and a first cross bar The mechanical linkage includes linear actuators that include piston-type actuators configured to move at least one payload in and out of a compartment. One end of each linear actuator is mounted to an individual rocking bar. The mechanical linkage includes at least one grapple, wherein each grapple comprises a hook mounted to a claw; and is mounted to a second cross bar. The hook is configured to open and close around the at least one payload; and rotate around the claw. The delivery mechanism includes a control circuit configured to open, close, and rotate the hook using a grapple motor and/or an end-of-travel switch; and to move the linear actuators relative to the rocking bars.

Abstract: Systems and methods for the application of surface normal calculations are illustrated. One embodiment includes a system for navigation, including: a processor; and instructions stored in a memory that when executed by the processor direct the processor. The processor obtains a set of sensor data, wherein sensor data includes a plurality of polarized images. The processor retrieves at least one navigation query; and a plurality of key-value pairs based on the polarized images. The processor inputs the at least one navigation query and the plurality of key-value pairs into a Cross-Attention Transformer that provides a set of weighted sums, wherein each weighted sum corresponds to: a certain key-value pair from the plurality of key-value pairs; and a certain sensor. The processor updates a model based on the set of weighted sums. The processor navigates the system within a 3D environment according, at least in part, to the model.

Abstract: Attention-free transformers are disclosed. Various implementations of attention-free transformers include a gating and pooling operation that allows the attention-free transformers to provide comparable or better results to those of a standard attention-based transformer, with improved efficiency and reduced computational complexity with respect to space and time.

Abstract: Systems and techniques performing autonomous navigation are illustrated. One embodiment includes a method for navigation. The method inputs a set of sensor data obtained from a plurality of sensors into at least one convolutional neural network (CNN). The at least one CNN generates a plurality of key-value pairs where each the key-value pair corresponds to an individual sensor from the plurality of sensors; and a value included in the key-value pair is determined based upon a subset of sensor data obtained from the individual sensor. The method inputs at least one navigation query and the plurality of key-value pairs into a Cross-Attention Transformer (CAT). The method obtains, from the CAT, a set of weighted sums, wherein each weighted sum corresponds to: a certain key-value pair; and a certain sensor from the plurality of sensors. The method updates a model depicting a 3D environment based on the set of weighted sums.

Abstract: Implementations of the subject technology relate to generative scene networks (GSNs) that are able to generate realistic scenes that can be rendered from a free moving camera at any location and orientation. A GSN may be implemented using a global generator and a locally conditioned radiance field. GSNs may employ a spatial latent representation as conditioning for a grid of locally conditioned radiance fields, and may be trained using an adversarial learning framework. Inverting a GSN may allow free navigation of a generated scene conditioned on one or more observations.

Abstract: A method includes receiving three-dimensional geometric elements as an input. The method also includes initializing geometric capsules by assigning one of the three-dimensional geometric elements to each of the geometric capsules and setting initial values for a pose component and a feature component of each of the geometric capsules. The method also includes one or more iterations of a routing procedure that includes assigning an additional one of the three-dimensional geometric elements to a respective one of the geometric capsules, based on correspondence of the additional one of the three-dimensional geometric elements to a surface defined based on the feature component of the respective one of the geometric capsules, and updating the feature component of each of the geometric capsules based on the three-dimensional geometric elements assigned to each of the geometric capsules. The method also includes outputting the geometric capsules including encoded three-dimensional data.

Abstract: A method includes defining a geometric capsule that is interpretable by a capsule neural network, wherein the geometric capsule includes a feature representation and a pose. The method also includes determining multiple viewpoints relative to the geometric capsule and determining a first appearance representation of the geometric capsule for each of the multiple viewpoints. The method also includes determining a transform for each of the multiple viewpoints that moves each of the multiple viewpoints to a respective transformed viewpoint and determining second appearance representations that each correspond to one of the transformed viewpoints. The method also includes combining the second appearance representations to define an agreed appearance representation. The method also includes updating the feature representation for the geometric capsule based on the agreed appearance representation.

Abstract: A system for training a neural network. A switch is linked to feature detectors in at least some of the layers of the neural network. For each training case, the switch randomly selectively disables each of the feature detectors in accordance with a preconfigured probability. The weights from each training case are then normalized for applying the neural network to test data.

Abstract: A method includes obtaining a two-dimensional image, obtaining a two-dimensional image annotation that indicates presence of an object in the two-dimensional image, obtaining three-dimensional sensor information, generating a top-down representation of the three-dimensional sensor information, and obtaining a top-down annotation that indicates presence of the object in the top-down representation. The method also includes determining a bottom surface of a three-dimensional cuboid based on map information, determining a position, a length, a width, and a yaw rotation of the three-dimensional cuboid based on the top-down annotation, and determining a height of the three-dimensional cuboid based on a two-dimensional image annotation, and the position, the length, the width, and the yaw rotation of the three-dimensional cuboid.

Abstract: At a centralized model trainer, one or more neural network based models are trained using an input data set. At least a first set of parameters of a model is transmitted to a model deployment destination. Using a second input data set, one or more adaptive parameters for the model are determined at the model deployment destination. Using the adaptive parameters, one or more inferences are generated at the model deployment destination.

Abstract: Sensor data captured by one or more sensors may be received at an analysis system. A neural network may be used to detect an object in the sensor data. A plurality of polygons surrounding the object may be generated in one or more subsets of the sensor data. A prediction of a future position of the object may be generated based at least in part on the polygons. One or more commands may be provided to a control system based on the prediction of the future position.

Abstract: Sensor data captured by one or more sensors may be received at an analysis system. A neural network may be used to detect an object in the sensor data. A plurality of polygons surrounding the object may be generated in one or more subsets of the sensor data. A prediction of a future position of the object may be generated based at least in part on the polygons. One or more commands may be provided to a control system based on the prediction of the future position.

Abstract: Implementations of the subject technology relate to generative scene networks (GSNs) that are able to generate realistic scenes that can be rendered from a free moving camera at any location and orientation. A GSN may be implemented using a global generator and a locally conditioned radiance field. GSNs may employ a spatial latent representation as conditioning for a grid of locally conditioned radiance fields, and may be trained using an adversarial learning framework. Inverting a GSN may allow free navigation of a generated scene conditioned on one or more observations.

Abstract: A method includes obtaining a two-dimensional image, obtaining a two-dimensional image annotation that indicates presence of an object in the two-dimensional image, determining a location proposal based on the two-dimensional image annotation, determining a classification for the object, determining an estimated size for the object based on the classification for the object, and defining a three-dimensional cuboid for the object based on the location proposal and the estimated size.

Abstract: Attention-free transformers are disclosed. Various implementations of attention-free transformers include a gating and pooling operation that allows the attention-free transformers to provide comparable or better results to those of a standard attention-based transformer, with improved efficiency and reduced computational complexity with respect to space and time.

Abstract: A method includes obtaining training samples that include images that depict objects and annotations of annotated key point locations for the objects. The method also includes training a machine learning model to determine estimated key point locations for the objects and key point uncertainty values for the estimated key point locations by minimizing a loss function that is based in part on a key point localization loss value that represents a difference between the annotated key point locations and the estimated key point locations values and is weighted by the key point uncertainty values.

Abstract: A system for training a neural network. A switch is linked to feature detectors in at least some of the layers of the neural network. For each training case, the switch randomly selectively disables each of the feature detectors in accordance with a preconfigured probability. The weights from each training case are then normalized for applying the neural network to test data.

Abstract: A system for training a neural network. A switch is linked to feature detectors in at least some of the layers of the neural network. For each training case, the switch randomly selectively disables each of the feature detectors in accordance with a preconfigured probability. The weights from each training case are then normalized for applying the neural network to test data.

Abstract: A method includes defining a geometric capsule that is interpretable by a capsule neural network, wherein the geometric capsule includes a feature representation and a pose. The method also includes determining multiple viewpoints relative to the geometric capsule and determining a first appearance representation of the geometric capsule for each of the multiple viewpoints. The method also includes determining a transform for each of the multiple viewpoints that moves each of the multiple viewpoints to a respective transformed viewpoint and determining second appearance representations that each correspond to one of the transformed viewpoints. The method also includes combining the second appearance representations to define an agreed appearance representation. The method also includes updating the feature representation for the geometric capsule based on the agreed appearance representation.

Abstract: A system employs an inspection neural network (INN) to inspect data generated during an inference process of a primary neural network (PNN) to generate an indication of reliability for an output generated by the PNN. The system includes a sensor configured to capture sensor data. Sensor data captured by the sensor is provided to a data analyzer to generate an output using the PNN. An analyzer inspector is configured to capture inspection data associated with the generation of the output by the data analyzer, and use the INN to generate an indication of reliability for the PNN's output based on the inspection data. The INN is trained using a set of training data that is distinct from the training data used to train the PNN.