Zoox, Inc.

Gowtham Garimella

Kai Zhenyu Wang

Marin Kobilarov

Andres Guillermo Morales Morales

Ethan Miller Pronovost

Samir Parikh

Jefferson Bradfield Packer

Andrew Scott Crego

Xiaosi Zeng

Joseph Funke

Chuang Wang

Gerrit Bagschik

Andraz Kavalar

Michael Haggblade

Kenneth Michael Siebert

Benjamin Isaac Mattinson

Linjun Zhang

William Anthony Silva

Gary Linscott

Timothy Caldwell

Mahsa Ghafarianzadeh

Rick Zhang

Rasmus Fonseca

Muhammad Farooq Rama

Ruikun Yu

Andreas Pasternak

Chunlei Dai

Yiteng Zhang

Aaron Huang

Yousef Ali Emam

Xiaobo Ren

Woodrow Zhouyuan Wang

A machine-learned architecture for determining whether an object is relevant to a vehicle's action planning may comprise a convolutional neural network, graph neural network, and/or multi-layer perceptron that may determine a relevance score associated with an object that indicates indicating whether an object is likely to impact operation(s) of a vehicle. In some examples, the machine-learned architecture may use scene information and/or an object track to determine the relevance score.

Machine-learned model for detecting object relevance to vehicle operation planning

Techniques are discussed for predicting an occupancy of visible region of an environment. For instance, a vehicle may generate sensor data representing an environment. The vehicle may then analyze the sensor data to determine an occluded region of the environment a visible region of the environment. Additionally, the vehicle may determine at least one prediction probability associated with occupancy of the visible region over a future period of time. In some instances, the vehicle determines the at least one prediction probability by inputting data representing at least the occluded region and the visible region into a machine learned model and receiving the at least one prediction probability from the machine learned model. Using the at least one prediction probability, the vehicle may then determine and perform one or more actions.

Predicting occupancy of objects in occluded regions

Techniques relating to determining regions based on intents of objects are described. In an example, a computing device onboard a first vehicle can receive sensor data associated with an environment of the first vehicle. The computing device can determine, based on the sensor data, a region associated with a second vehicle proximate the first vehicle that is to be occupied by the second vehicle while the vehicle performs a maneuver. Further, the computing device can determine an instruction for controlling the first vehicle based at least in part on the region.

Spatial prediction

Techniques for improving operational decisions of an autonomous vehicle are discussed herein. In some cases, a system may generate reference graphs associated with a route of the autonomous vehicle. Such reference graphs can comprise precomputed feature vectors based on grid regions and/or lane segments. The feature vectors are usable to determine scene context data associated with static objects to reduce computational expenses and compute time.

System for generating scene context data using a reference graph

Techniques are discussed herein for generating and using graph neural networks (GNNs) including vectorized representations of map elements and entities within the environment of an autonomous vehicle. Various techniques may include vectorizing map data into representations of map elements, and object data representing entities in the environment of the autonomous vehicle. In some examples, the autonomous vehicle may generate and/or use a GNN representing the environment, including nodes stored as vectorized representations of map elements and entities, and edge features including the relative position and relative yaw between the objects. Machine-learning inference operations may be executed on the GNN, and the node and edge data may be extracted and decoded to predict future states of the entities in the environment.

Graph neural networks with vectorized object representations in autonomous vehicle systems

Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a distribution of predicted positions for the object in the future that meet a criterion, allowing for more efficient sampling. A predicted position of the object in the future may be determined by sampling from the distribution.

Focusing prediction distribution output for efficient sampling

Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a distribution of predicted positions for the object in the future. A predicted position of the object at a subsequent timestep may be determined by sampling from the distribution of predicted positions according to various sampling strategies. Alternatively, the predicted position of the object may be overwritten using a candidate position of the object.

Prediction sampling techniques

Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a predicted position of the object at a subsequent timestep. Further, a predicted trajectory of the object may be determined using predicted positions of the object at various timesteps.

Encoding relative object information into node edge features

Techniques are discussed for interaction probabilities associated with regions of an environment around a vehicle. An interaction probability of a region may indicate a likelihood an object positioned at the region will interact with the vehicle. A top-down multi-channel image representing a top-down view of the environment and objects therein may be generated and input to a machine learned (ML) model. The ML model may output a probability map, a portion of the probability map comprising a region and an interaction probability associated with the region that indicates a likelihood objects positioned at the region will interact with the vehicle. A priority for resource assignment or analysis may be determined based on the interaction probability for an object positioned in the region. Control of the vehicle may be performed based at least in part on the priority for resource assignment or analysis.

Machine learned interaction prediction from top-down representation

Model-based control of dynamical systems typically requires accurate domain-specific knowledge and specifications system components. Generally, steering actuator dynamics can be difficult to model due to, for example, an integrated power steering control module, proprietary black box controls, etc. Further, it is difficult to capture the complex interplay of non-linear interactions, such as power steering, tire forces, etc. with sufficient accuracy. To overcome this limitation, a recurring neural network can be employed to model the steering dynamics of an autonomous vehicle. The resulting model can be used to generate feedforward steering commands for embedded control. Such a neural network model can be automatically generated with less domain-specific knowledge, can predict steering dynamics more accurately, and perform comparably to a high-fidelity first principle model when used for controlling the steering system of a self-driving vehicle.

Steering control for vehicles

A reward determined as part of a machine learning technique, such as reinforcement learning, may be used to control an adversarial agent in a simulation such that a component for controlling motion of the adversarial agent is trained to reduce the reward. Training the adversarial agent component may be subject to one or more constraints and/or may be balanced against one or more additional goals. Additionally or alternatively, the reward may be used to alter scenario data so that the scenario data reduces the reward, allowing the discovery of difficult scenarios and/or prospective events.

Adversarial agent controls generation and problematic scenario forecasting

Techniques for top-down scene discrimination are discussed. A system receives scene data associated with an environment proximate a vehicle. The scene data is input to a convolutional neural network (CNN) discriminator trained using a generator and a classification of the output of the CNN discriminator. The CNN discriminator generates an indication of whether the scene data is a generated scene or a captured scene. If the scene data is data generated scene, the system generates a caution notification indicating that a current environmental situation is different from any previous situations. Additionally, the caution notification is communicated to at least one of a vehicle system or a remote vehicle monitoring system.

Top-down scene discrimination

Techniques for modeling the probability distribution of errors in perception systems are discussed herein. For example, techniques may include modeling error distribution for attributes such as position, size, pose, and velocity of objects detected in an environment, and training a mixture model to output specific error probability distributions based on input features such as object classification, distance to the object, and occlusion. The output of the trained model may be used to control the operation of a vehicle in an environment, generate simulations, perform collision probability analyses, and to mine log data to detect collision risks.

Perception error modeling

Techniques for top-down scene generation are discussed. A generator component may receive multi-dimensional input data associated with an environment. The generator component may generate, based at least in part on the multi-dimensional input data, a generated top-down scene. A discriminator component receives the generated top-down scene and a real top-down scene. The discriminator component generates binary classification data indicating whether an individual scene in the scene data is classified as generated or classified as real. The binary classification data is provided as a loss to the generator component and the discriminator component.

Top-down scene generation

Techniques for determining predictions on a top-down representation of an environment based on object movement are discussed herein. Sensors of a first vehicle (such as an autonomous vehicle) may capture sensor data of an environment, which may include object(s) separate from the first vehicle (e.g., a vehicle, a pedestrian, a bicycle). A multi-channel image representing a top-down view of the object(s) and the environment may be generated based in part on the sensor data. Environmental data (object extents, velocities, lane positions, crosswalks, etc.) may also be encoded in the image. Multiple images may be generated representing the environment over time and input into a prediction system configured to output a trajectory template (e.g., general intent for future movement) and a predicted trajectory (e.g., more accurate predicted movement) associated with each object. The prediction system may include a machine learned model configured to output the trajectory template(s) and the predicted trajector(ies).

Growing community of inventors

Baltimore, MD, United States of America

Gowtham Garimella