Zoox, Inc.

Lyft, Inc.

Woven Planet North America, Inc.

Woven by Toyota, U.s., Inc.

Ashesh Jain

Lei Zhang

Dragomir Dimitrov Anguelov

Yunjian Jiang

Mushfiqur Rouf

Henru Wang

Danfei Xu

Michael Jared Benisch

Taggart C Matthiesen

Meng Gao

Anantha Rao Kancherla

In one embodiment, a computing system of a vehicle may access sensor data associated with a surrounding environment of a vehicle. The system may generate, based on the sensor data, a first trajectory having one or more first driving characteristics for navigating the vehicle in the surrounding environment. The system may generate a second trajectory having one or more second driving characteristics by modifying the one or more first driving characteristics of the first trajectory. The modifying may use adjustment parameters based on one or more human-driving characteristics of observed human-driven trajectories such that the one or more second driving characteristics satisfy a similarity threshold relative to the one or more human-driving characteristics. The system may determine, based on the second trajectory, vehicle operations to navigate the vehicle in the surrounding environment.

Using machine learning models for generating human-like trajectories

Examples disclosed herein may involve a computing system that is configured to (i) obtain previously-derived perception data for a collection of sensor data including a sequence of frames observed by a vehicle within one or more scenes, where the previously-derived perception data includes a respective set of object-level information for each of a plurality of objects detected within the sequence of frames, (ii) derive supplemental object-level information for at least one object detected within the sequence of frames that adds to the previously-derived object-level information for the at least one object, (iii) augment the previously-derived perception data to include the supplemental object-level information for the at least one object, and (iv) store the augmented perception data in an arrangement that encodes a hierarchical relationship between the plurality of objects, the sequence of frames, and the one or more scenes.

Systems and methods for augmenting perception data with supplemental information

In one embodiment, a method includes, by a computing system associated with a vehicle, determining a current location of the vehicle in a first region, identifying one or more first sets of model parameters associated with the first region and one or more second sets of model parameters associated with a second region, generating, using one or more machine-learning models based on the first sets of model parameters, one or more first inferences based on first sensor data captured by the vehicle, switching the configurations of the models from the first sets of model parameters to the second sets of model parameters, generating, using the models having configurations based on the second sets of model parameters, one or more second inferences based on second sensor data generated by the sensors of the vehicle in the second region, and causing the vehicle to perform one or more operations based on the second inferences.

Geolocalized models for perception, prediction, or planning

Systems, methods, and non-transitory computer-readable media can determine first sensor data captured by a first sensor of a vehicle. Second sensor data captured by a second sensor of the vehicle can be determined. Information describing the first sensor data and the second sensor data can be provided to a machine learning model trained to predict whether a pair of sensors are calibrated or mis-calibrated based on sensor data captured by the pair of sensors. A determination is made whether the first sensor and the second sensor are calibrated or mis-calibrated based on an output from the machine learning model.

Approaches for determining sensor calibration

A three-dimensional bounding box is determined from a two-dimensional image and a point cloud. A feature vector associated with the image and a feature vector associated with the point cloud may be passed through a neural network to determine parameters of the three-dimensional bounding box. Feature vectors associated with each of the points in the point cloud may also be determined and considered to produce estimates of the three-dimensional bounding box on a per-point basis.

Three-dimensional bounding box from two-dimensional image and point cloud data

Systems, methods, and non-transitory computer-readable media can determine sensor data captured by at least one sensor of a vehicle over a set of time intervals while navigating an environment. Three-dimensional data describing the environment over the set of time intervals can be determined from the captured sensor data. The three-dimensional data capturing a traffic motion pattern for at least one direction of travel. Image data of at least one traffic light in the environment can be determined over the set of time intervals from the captured sensor data. A state of the at least one traffic light can be predicted based at least in part on the three-dimensional data describing the environment and the image data of at least one traffic light in the environment over the set of time intervals.

Growing community of inventors

Palo Alto, CA, United States of America

Ashesh Jain