Patent Application
20200387155
On some occasions, such as when a vehicle that has autonomous driving capabilities (an AV) is driving on a road and experiences an event—such as system faults, extreme weather conditions, and temporary detours—it may be useful to have a remotely located person provide assistance.
In some implementations, the technologies described in this document include a teleoperation system that interacts with an AV system to handle various types of events, some of which may induce risks (e.g., collisions, traffic jams and damages) or may prohibit or inhibit the AV that is part of the AV system from traveling along a planned trajectory. In some examples, to handle the events, the AV system may communicate with a teleoperation system where a teleoperator provides teleoperations to the AV system.
In some cases, the teleoperation system may comprise a client onboard the AV or associated with the AV system and a server remote to the AV or the AV system. In some cases, the client and the server are both onboard the AV. In some applications, the server and the client may be separated into two different computing devices; in some cases, they may be integrated into a single computing device.
In some implementations, a function or a step described as a part of a teleoperation client may be realized as a part of a teleoperation server. Similarly, a function or a step described as a part of a teleoperation server may be realized as a part of a teleoperation client. In some cases, a function or a step can be a part of a teleoperation server and a part of a teleoperation client.
In general, in one aspect, a method comprises: (a) determining that intervention in an operation of one or more autonomous driving capabilities of a vehicle is appropriate; (b) based on the determination, enabling a person to provide information for an intervention; and (c) causing the intervention in the operation of the one or more autonomous driving capabilities of the vehicle. Determining that intervention is appropriate may comprise receiving a request for intervention. Determining that intervention is appropriate may comprise receiving information about a status or environment of the vehicle or a related AV system. The status or the environment of the vehicle may comprise a functionality of a hardware component or software of the vehicle or the AV system.
In some implementations, the information about the status or the environment of the vehicle or the AV system may comprise a signal from a hardware component or software of the vehicle or the AV system. Determining that an intervention is appropriate may comprise analyzing the signal. Analyzing the signal may comprise detecting unexpected data or absence of expected data. Analyzing the signal may comprise evaluating a mismatch between a measured quantity and a model-estimated quantity for the hardware component or software. Analyzing the signal may comprise using pattern recognition to evaluate an abnormal pattern in the signal. The abnormal pattern may be learned by a machine learning algorithm. Analyzing the signal may comprise inferring a malfunction in the hardware component or the software. Analyzing the signal may comprise detecting an unknown object present in the environment of the vehicle. Analyzing the signal may comprise inferring an event that is or will be happening in the environment of the vehicle.
In some implementations, the request may comprise a request initiated by a remote operator. The request may comprise data associated with status or environment of a vehicle or a related AV system. The request may comprise one or more signals from one or more hardware components of the vehicle or a related AV system. The request may comprise one or more signals from one or more software processes of the vehicle.
In some implementations, the method may comprise, based on the determination, causing a fallback intervention in the operation of the one or more autonomous driving capabilities of the vehicle. The fallback intervention may comprise causing the vehicle or a related AV system to enter a fully autonomous driving mode, a semi-autonomous driving mode, or a fully manual driving mode. The fallback intervention may comprise causing the vehicle to operate at a reduced velocity. The fallback intervention may comprise identifying a safe-to-stop location. The fallback intervention may comprise generating a new trajectory to the safe-to-stop location. The fallback intervention may comprise invoking a backup hardware component or a backup software process. The fallback intervention may comprise evaluating functional hardware components or software processes required to operate the vehicle.
In some implementations, determining that intervention is appropriate may comprise evaluating one or more active events associated with the vehicle or a related AV system, or the environment of the vehicle. Evaluating one or more events may comprise merging two or more active events. Enabling the person to provide information for an intervention may comprise maintaining a queue based on one or more determinations that intervention is appropriate. Maintaining the queue may comprise prioritizing an intervention based on one or more of the following: a decision tree, a combinatorial optimization, a machine algorithm, and a past intervention. Enabling the person to provide information for an intervention may comprise allocating the person to provide the information based on availability of the person, and one or more of: (a) time, (b) knowledge of the vehicle, (c) knowledge of the environment of the vehicle, or (d) a language.
In some implementations, enabling the person to provide information for an intervention may comprise presenting an interactive interface to the person. Presenting an interactive interface may comprise presenting a field of view or a bird's-eye of a vision sensor of the vehicle. Presenting an interactive interface may comprise presenting current or past or both perception information. Presenting an interactive interface may comprise presenting current or past or both trajectories. Presenting an interactive interface may comprise presenting current or past or both motion planning information. Presenting an interactive interface may comprise presenting a system diagram of the vehicle, the system diagram comprising one or more hardware components, or one or more software processes, or both.
In some implementations, the information for the intervention may comprise a current location of the vehicle determined by the person. The intervention may comprise treating the current location identified by the person as prior knowledge and using an inference algorithm to update the current location. The intervention may be based on the person identifying a goal location of the vehicle. The intervention may comprise treating the goal location identified by the person as prior knowledge and using an inference algorithm to update the goal location.
In some implementations, the method may include an intervention comprising a trajectory to be found by the person. The intervention may comprise treating the trajectory identified by the person as prior knowledge and using an inference algorithm to update the trajectory. The intervention may comprise one or more trajectory sampling points identified by the person. The intervention may comprise inferring a trajectory or a trajectory segment based on the one or more trajectory sampling points. Inferring a trajectory or a trajectory segment may be based on one or more trajectory primitives. The intervention may comprise concatenating two trajectory segments. Concatenating two trajectory segments may comprise smoothing the trajectory segments and smoothing a speed profile across the trajectory segments. The intervention may comprise specifying one or more un-traversable road segments. The intervention may comprise setting a speed profile. The intervention may comprise treating the speed profile as prior knowledge and using an inference algorithm to update the speed profile.
In some implementations, the intervention may be based on inferring a speed profile by a learning algorithm. The intervention may be based on inferring a steering angle by a learning algorithm. The intervention may comprise enabling, editing or disabling a hardware component or a software process. The intervention may comprise enabling, editing or disabling a subcomponent of a hardware component or a processing step of a software process.
In some implementations, the method may include an intervention comprising overwriting a travel preference or a travel rule. In some implementations, the method may include an intervention comprising editing data, the data comprising one or more of the following: a map, sensor data in the vehicle or a related AV system, trajectory data in the vehicle or a related AV system, vision data in the vehicle or a related AV system, or any past data in the vehicle or a related AV system.
In some implementations, the method may include configuring the vehicle or a related AV system based on a command. Configuring the vehicle or the AV system based on a command may comprise treating the command as prior knowledge and using an inference algorithm to update the command. A command may comprise one or more of the following: a trajectory, a label, a process control, an annotation, and a machine instruction.
In general, in an aspect, a method comprises (a) receiving an intervention request regarding an operation of one or more autonomous driving capabilities of a vehicle; (b) causing a person to interact with the vehicle over a communication channel; and (c) issuing an intervention to configure the operation of the one or more autonomous driving capabilities of a vehicle.
In some implementations, the method may comprise receiving or generating or analyzing information about a status or environment of the vehicle. The information about the status or the environment of the vehicle may comprise a functionality of a hardware component or software of the vehicle. The information about the status or the environment of the vehicle may comprise a signal from a hardware component or software of the vehicle. The information about the status or the environment of the vehicle may comprise presence of unexpected data or absence of expected data. The information about the status or the environment of the vehicle may comprise a mismatch between a measured quantity and a model-estimated quantity for a hardware component or software of the vehicle.
In some implementations, analyzing the information may comprise using pattern recognition to evaluate an abnormal pattern in the information. The abnormal pattern may be learned by a machine learning algorithm. Analyzing the information may comprise inferring a malfunction in the hardware component or the software. Analyzing the information may comprise detecting an unknown object present in the environment of the vehicle. Analyzing the information may comprise inferring an event that is or will be happening in the environment of the vehicle.
In some implementations, the intervention request may comprise a request initiated by the person or a second person. The intervention request may comprise data associated with status or environment of a vehicle or a related AV system. The intervention request may comprise one or more signals from one or more hardware components of the vehicle or a related AV system. The intervention request may comprise one or more signals from one or more software processes of the vehicle.
In some implementations, determining that intervention is appropriate may comprise evaluating one or more active events associated with the vehicle or a related AV system, or the environment of the vehicle. Evaluating one or more events may comprise merging two or more active events.
In some implementations, the method may comprise maintaining a queue of one or more intervention requests. Maintaining the queue may comprise prioritizing an intervention based on one or more of the following: a decision tree, a combinatorial optimization, a machine algorithm, and a past intervention.
In some implementations, the method may comprise allocating the person to interact with the vehicle based on availability of the person, and one or more of: (a) time, (b) knowledge of the vehicle, (c) knowledge of the environment of the vehicle, or (d) a language.
In some implementations, the method may comprise presenting an interactive interface to the person. Presenting an interactive interface may comprise presenting a field of view or a bird's-eye of a vision sensor of the vehicle. Presenting an interactive interface may comprise presenting current or past, or both, perception information. Presenting an interactive interface may comprise presenting current or past, or both, trajectories. Presenting an interactive interface may comprise presenting current or past, or both, motion planning information. Presenting an interactive interface comprises presenting a system diagram of the vehicle, the system diagram comprising one or more hardware components, or one or more software processes, or both.
In some implementations, the method may include an intervention comprising a current location of the vehicle determined by the person; the intervention may comprise treating the current location identified by the person as prior knowledge and using an inference algorithm to update the current location. The intervention may comprise identifying a goal location for the vehicle; the intervention may comprise treating the goal location identified as prior knowledge and using an inference algorithm to update the goal location. The intervention may comprise a trajectory to be found by the person; the intervention may comprise treating the trajectory identified by the person as prior knowledge and using an inference algorithm to update the trajectory. The intervention may comprise one or more trajectory sampling points identified by the person; the intervention may comprise inferring a trajectory or a trajectory segment based on the one or more trajectory sampling points. Inferring a trajectory or a trajectory segment may be based on one or more trajectory primitives. The intervention may comprise concatenating two trajectory segments. Concatenating two trajectory segments may comprise smoothing the trajectory segments and smoothing a speed profile across the trajectory segments. In some implementations, the intervention may comprise specifying one or more un-traversable road segments.
In some implementations, the intervention may comprise setting a speed profile; the intervention may comprise treating the speed profile as prior knowledge and using an inference algorithm to update the speed profile. The intervention may be based on inferring a speed profile by a learning algorithm. The intervention may be based on inferring a steering angle by a learning algorithm. The intervention may comprise enabling, editing or disabling a hardware component or a software process. The intervention may comprise enabling, editing or disabling a subcomponent of a hardware component or a processing step of a software process.
In some implementations, the intervention may comprise overwriting a travel preference or a travel rule. The intervention may comprise editing data, the data comprising one or more of the following: a map, sensor data in the vehicle, trajectory data in the vehicle, vision data in the vehicle, or any past data in the vehicle. Configuring the operation of the one or more autonomous driving capabilities may comprise treating an intervention as prior knowledge and using an inference algorithm to update the intervention for a purpose of the configuration. An intervention may comprise one or more of the following: a trajectory, a label, a process control, an annotation, and a machine instruction.
In general, in an aspect, implementations include a vehicle with autonomous driving capabilities comprising, and the vehicle may comprise (a) steering, acceleration, and deceleration devices that respond to controlling signals from a driving control system to drive the vehicle autonomously on a road network; (b) a monitoring element on the vehicle that generates an intervention request for the vehicle to engage in an intervention with a person; and (c) a communication element that receives a command from the person to the driving control system for the steering, acceleration, and deceleration devices to cause the vehicle to maneuver to a goal location.
In some implementations, the vehicle may comprise a processor that receives information about a status or environment of the vehicle to determine that the intervention is appropriate. The status or the environment of the vehicle may comprise a functionality of a hardware component or software of the vehicle. The information about the status or the environment of the vehicle may comprise a signal from a hardware component or software of the vehicle.
In some implementations, the vehicle may include determining that intervention is appropriate by analyzing the signal. Analyzing the signal may comprise detecting unexpected data or absence of expected data. Analyzing the signal may comprise evaluating a mismatch between a measured quantity and a model-estimated quantity for the hardware component or software. Analyzing the signal may comprise using pattern recognition to evaluate an abnormal pattern in the signal. An abnormal pattern may be learned from a machine learning algorithm. Analyzing the signal may comprise inferring a malfunction in the hardware component or the software. Analyzing the signal may comprise detecting an unknown object present in the environment of the vehicle. Analyzing the signal may comprise inferring an event that is or will be happening in the environment of the vehicle.
In some implementations, the request may comprise a request initiated by a remote operator. The request may comprise data associated with status or environment of a vehicle. The request may comprise one or more signals from one or more hardware components of the vehicle. The request may comprise one or more signals from one or more software processes of the vehicle.
In some implementations, the vehicle may comprise a processor causing a fallback intervention in the driving control system. The fallback intervention may comprise causing the vehicle to enter a fully autonomous driving mode, a semi-autonomous driving mode, or a fully manual driving mode. The fallback intervention may comprise causing the vehicle to operate at a reduced velocity. The fallback intervention may comprise identifying a safe-to-stop location. The fallback intervention may comprise generating a new trajectory to the safe-to-stop location. The fallback intervention may comprise invoking a backup hardware component or a backup software process. The fallback intervention may comprise evaluating functional hardware components or software processes required to operate the vehicle.
In some implementations, the vehicle may comprise a processor evaluating one or more active events associated with the vehicle, or the environment of the vehicle. Evaluating one or more active events may comprise merging two or more active events.
In some implementations, the vehicle may comprise a processor enabling the person to provide information for an intervention comprises maintaining a queue based on one or more determinations that intervention is appropriate. Maintaining the queue may comprise prioritizing an intervention based on one or more of the following: a decision tree, a combinatorial optimization, a machine algorithm, and a past intervention. Enabling the person to provide information for an intervention may comprise allocating the person to provide the information based on availability of the person, and one or more of: (a) time, (b) knowledge of the vehicle, (c) knowledge of the environment of the vehicle, or (d) a language. Enabling the person to provide information for an intervention may comprise presenting an interactive interface to the person. Presenting an interactive interface may comprise presenting a field of view or a bird's-eye of a vision sensor of the vehicle. Presenting an interactive interface may comprise presenting current or past, or both, perception information. Presenting an interactive interface may comprise presenting current or past, or both, trajectories. Presenting an interactive interface may comprise presenting current or past, or both, motion planning information. Presenting an interactive interface may comprise presenting a system diagram of the vehicle, the system diagram comprising one or more hardware components, or one or more software processes, or both.
In some implementations, the intervention may comprise a current location of the vehicle determined by the person. The intervention may comprise treating the current location identified by the person as prior knowledge and using an inference algorithm to update the current location.
In some implementations, the intervention may be based on the person identifying a goal location of the vehicle. The intervention may comprise treating the goal location identified by the person as prior knowledge and using an inference algorithm to update the goal location.
In some implementations, the intervention may comprise a trajectory to be found by the person. The intervention may comprise treating the trajectory identified by the person as prior knowledge and using an inference algorithm to update the trajectory.
In some implementations, the intervention may comprise one or more trajectory sampling points identified by the person. The intervention may comprise inferring a trajectory or a trajectory segment based on the one or more trajectory sampling points. Inferring a trajectory or a trajectory segment may be based on one or more trajectory primitives. The intervention may comprise concatenating two trajectory segments. Concatenating two trajectory segments may comprise smoothing the trajectory segments and smoothing a speed profile across the trajectory segments.
In some implementations, an intervention may comprise specifying one or more un-traversable road segments. An intervention may comprise setting a speed profile. An intervention may comprise treating the speed profile as prior knowledge and using an inference algorithm to update the speed profile. An intervention may be based on inferring a speed profile by a learning algorithm. An intervention may be based on inferring a steering angle by a learning algorithm. An intervention may comprise enabling, editing or disabling a hardware component or a software process. An intervention may comprise enabling, editing or disabling a subcomponent of a hardware component or a processing step of a software process. An intervention may comprise overwriting a travel preference or a travel rule. An intervention may comprise editing data, the data comprising one or more of the following: map, sensor data, trajectory data, vision data, or any past data.
In some implementations, the vehicle may comprise a processor configuring the vehicle or a related AV system based on a command. Configuring the vehicle or the AV system based on a command may comprise treating the command as prior knowledge and using an inference algorithm to update the command. A command may comprise one or more of the following: a trajectory, a label, a process control, an annotation, and a machine instruction.
In another aspect, implementations include an apparatus comprising: (a) a processor configured to (1) receive an intervention request regarding operation of a vehicle and (2) extract motion information or perception information from the intervention request, and (b) a display configured to (1) display the motion information or the perception information and (2) allow a user to interact with operation of the vehicle.
In some implementations, an intervention request may comprise a request initiated by a remote operator. An intervention request may comprise data associated with status or environment of the vehicle or a related AV system. An intervention request may comprise one or more signals from one or more hardware components of the vehicle or a related AV system. An intervention request may comprise one or more signals from one or more software processes of the vehicle.
In some implementations, the display may be configured to present an interactive interface comprising a field of view or a bird's-eye of a vision sensor of the vehicle. The display may be configured to present an interactive interface comprising current or past or both perception information. The display may be configured to present an interactive interface comprising current or past or both trajectories. The display may be configured to present an interactive interface comprising current or past or both motion planning information. The display may be configured to present an interactive interface comprising a system diagram of the vehicle, the system diagram comprising one or more hardware components, or one or more software processes, or both.
In some implementations, the apparatus may include a processor that converts one or more interactions from the user into an intervention for the operation of the vehicle. One or more interactions may comprise a current location of the vehicle determined by the user; a processor may treat the current location identified by the user as prior knowledge and uses an inference algorithm to generate an updated current location as an intervention. One or more interactions may comprise a goal location of the vehicle identified by the user; a processor may treat the goal location as prior knowledge and uses an inference algorithm to generate an updated goal location as an intervention. One or more interactions may comprise a trajectory identified by the user; a processor may treat the trajectory as prior knowledge and uses an inference algorithm to generate an updated trajectory as an intervention. One or more interactions comprise one or more trajectory sampling points identified by the person; a processor may infer a trajectory or a trajectory segment based on the one or more trajectory sampling points. A processor may infer a trajectory or a trajectory segment based on one or more trajectory primitives. A processor may concatenate two trajectory segments. Concatenating two trajectory segments may comprise smoothing the trajectory segments and smoothing a speed profile across the trajectory segments. One or more interactions may comprise specifying one or more un-traversable road segments. One or more interactions may comprise setting a speed profile. A processor may treat the speed profile as prior knowledge and use an inference algorithm to generate an updated speed profile as an intervention. A processor may infer a speed profile by a learning algorithm, and the speed profiles may be included in the intervention. A processor may infer a steering angle by a learning algorithm, and the steering angle may be included in the intervention. One or more interactions or an intervention may comprise enabling, editing or disabling a hardware component or a software process. One or more interactions or an intervention comprises enabling, editing or disabling a subcomponent of a hardware component or a step of a software process. One or more interactions or an intervention may comprise overwriting a travel preference or a travel rule. One or more interactions or an intervention may comprise editing data, the data comprising one or more of the following: a map, sensor data in the vehicle or a related AV system, trajectory data in the vehicle or a related AV system, vision data in the vehicle or a related AV system, or any past data in the vehicle or a related AV system. An intervention may comprise one or more of the following: a trajectory, a label, a process control, an annotation, and a machine instruction.
In general, in an aspect, a method comprises: (a) causing a vehicle to drive in an autonomous mode on a road, the vehicle comprising one or more autonomous driving capabilities; (b) receiving an intervention regarding an operation of the one or more autonomous driving capabilities; and (c) analyzing the intervention and configuring one or more hardware components or one or more software processes of the vehicle.
In some implementations, an intervention may comprise a current location of the vehicle; analyzing the intervention may comprise treating the current location in the intervention as prior knowledge and using an inference algorithm to update the current location. An intervention may comprise a goal location; analyzing the intervention may comprise treating the goal location in the intervention as prior knowledge and using an inference algorithm to update the goal location. An intervention may comprise a trajectory; analyzing the intervention may comprise treating the trajectory in the intervention as prior knowledge and using an inference algorithm to update the trajectory. An intervention may comprise one or more trajectory sampling points; analyzing the intervention may comprise treating the one or more trajectory sampling points as prior knowledge and using an inference algorithm to update the one or more trajectory sampling points. Analyzing an intervention may comprise inferring a trajectory or a trajectory segment based on the one or more trajectory sampling points. Inferring a trajectory or a trajectory segment may be based on one or more trajectory primitives. Inferring a trajectory or a trajectory segment may comprise concatenating two trajectory segments. Concatenating two trajectory segments may comprise smoothing the trajectory segments and smoothing a speed profile across the trajectory segments. An intervention may comprise specifying one or more un-traversable road segments. An intervention may comprise inferring or setting a speed profile. Analyzing an intervention may comprise treating the speed profile as prior knowledge and using an inference algorithm to update the speed profile. Analyzing an intervention may comprise inferring a speed profile by a learning algorithm. Analyzing an intervention may comprise inferring a steering angle by a learning algorithm. Analyzing an intervention may comprise enabling, editing or disabling a hardware component or a software process. Analyzing an intervention may comprise enabling, editing or disabling a subcomponent of a hardware component or a processing step of a software process. An intervention may comprise overwriting a travel preference or a travel rule. An intervention may comprise editing data, the data comprising one or more of the following: a map, sensor data in the vehicle or a related AV system, trajectory data in the vehicle or a related AV system, vision data in the vehicle or a related AV system, or any past data in the vehicle or a related AV system.
In general, in an aspect, a method comprises: (a) receiving from a remote operator machine-readable instructions regarding an operation of a vehicle; and (b) configuring the vehicle to execute the machine-readable instructions. The vehicle may comprise one or more autonomous driving capabilities. Machine-readable instructions may represent one or more of the following: a current location, a goal location, one or more trajectories, one or more trajectory sampling points, one or more speed profiles, or one or more un-traversable road segments. Machine-readable instructions may comprise enabling, editing or disabling a hardware component or a software process. Machine-readable instructions may comprise enabling, editing or disabling a subcomponent of a hardware component or a processing step of a software process. Machine-readable instructions may comprise overwriting a travel preference or a travel rule. Machine-readable instructions may comprise editing data, the data comprising one or more of the following: a map, sensor data in the vehicle or a related AV system, trajectory data in the vehicle or a related AV system, vision data in the vehicle or a related AV system, or any past data in the vehicle or a related AV system.
These and other aspects, features, and implementations can be expressed as methods, apparatus, systems, components, program products, methods of doing business, means or steps for performing a function, and in other ways.
These and other aspects, features, and implementations will become apparent from the following descriptions, including the claims.
The term “autonomous driving capability” is used broadly to include, for example, any function, feature, or facility that can participate in the driving of an AV other than by a person manipulating a steering wheel, accelerator, brake, or other physical controller of the AV.
The term “teleoperation” is used broadly to include, for example, any instruction, guidance, command, request, order, directive, or other control of or interaction with an autonomous driving capability of an AV, sent to the AV or the AV system by a communication channel (e.g., wireless or wired). This document sometimes uses the term “teleoperation command” interchangeably with “teleoperation.” Teleoperations are examples of interventions.
The term “teleoperator” is used broadly to include, for example, any person or any software process or hardware device or any combination of them that initiates, causes, or is otherwise the source of a teleoperation. A teleoperator may be local to the AV or AV system (e.g., occupying the AV, standing next to the AV, or one or more steps away from the AV), or remote from the AV or AV system (e.g., at least 1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100, 200, 300, 400, 500, 600, 700, 900, or 1000 meters away from the AV).
The term “teleoperation event” is used broadly to include, for example, any occurrence, act, circumstance, incident, or other situation for which a teleoperation would be appropriate, useful, desirable, or necessary.
The term “teleoperation request” is used broadly to include, for example, any communication from an AV or an AV system to a teleoperator or other part of a teleoperation system in connection with a teleoperation.
The term “tele-interact” or “tele-interaction” is used broadly to include, for example, any virtual interaction between a teleoperator and a hardware component or a software process of an AV or an AV system.
The term “fallback operation” is used broadly to include, for example, any fashion, form, or method of action, performance, or activity of an autonomous driving capability of an AV after a teleoperation request and before or while a corresponding teleoperation is received and executed by the AV system.
The term “trajectory” is used broadly to include, for example, any path or route from one place to another; for instance, a path from a pickup location to a drop off location.
The term “goal” or “goal position” is used broadly to include, for example, a place to be reached by an AV, including, for example, an interim drop off location, a final drop off location, or a destination, among others.
This document describes technologies applicable to any vehicles that have one or more autonomous driving capabilities including fully autonomous vehicles, highly autonomous vehicles, and conditionally autonomous vehicles, such as so-called Level 5, Level 4 and Level 3 vehicles, respectively (see SAE International's standard J3016: Taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems, which is incorporated by reference in its entirety, for more details on the classification of levels of autonomy in vehicles). Vehicles with autonomous driving capabilities may attempt to control the steering or speed of the vehicles. The technologies descried in this document can be applied to partially autonomous vehicles and driver assisted vehicles, such as so called Level 2 and Level 1 vehicles (see SAE International's standard J3016: Taxonomy and Definitions for Terms Related to On-Road Motor Vehicle Automated Driving Systems). One or more of the Level 1, 2, 3, 4 and 5 vehicle systems may automate certain vehicle operations (e.g., steering, braking, and using maps) under certain driving conditions based on analysis of sensor inputs. The technologies described in this document can benefit vehicles in any levels, ranging from fully autonomous vehicles to human-operated vehicles.
As shown in
The driving of an AV typically is supported by an array of technologies 18 and 20, (e.g., hardware, software, and stored and real time data) that this document together (and with the AV 10) refers to as an AV system 22. In some implementations, one or some or all of the technologies are onboard the AV. In some cases, one or some or all of the technologies are at another location such as at a server (e.g., in a cloud computing infrastructure). Components of an AV system can include one or more or all of the following (among others).
A teleoperation system, which may be remote or local or a combination of them to the AV or AV system, can enable a teleoperator to interact with the AV system (e.g., providing commands, visualizing a driving condition, and investigating functionality of a hardware component or software process) via a communication channel. The interactions may assist the AV system to adequately respond to various events.
Referring to
In step 304, the teleoperator accepts the teleoperation request and engages in the tele-interaction. The tele-interactions can vary; for example, the teleoperation server may recommend possible teleoperations through an interface to the teleoperator, and the teleoperator can select one or more of the recommended teleoperations and cause the teleoperations to be sent to the AV system. In some implementations, the teleoperation server renders an environment of the AV system through a user interface to the teleoperator, and the teleoperator can see the environment to select an optimal teleoperation. In some cases, the teleoperator may enter computer codes as a teleoperation. In some examples, the teleoperator uses the interface to draw a recommended trajectory for the AV along which to continue its driving.
Based on the tele-interaction, the teleoperator may issue a suitable teleoperation, which is then processed by a teleoperation handling process (336 in
The AV system monitoring process 420 may receive system information and data 412 to monitor the operation status (e.g., velocity, acceleration, steering, data communications, perception, and trajectory planning) of the AV system 410. The operation status may be based on directly reading outputs of hardware components or software processes or both of the AV system 410, or indirectly inferring, e.g., computationally or statistically, the outputs by measuring associated quantities, or both. In some implementations, the AV system monitoring process 420 may derive information (e.g., computing a statistic, or comparing monitored conditions with knowledge in a database) from the operation status. Based on the monitored operation status or derived information or both, the monitoring process 420 may determine a teleoperation event 422 for which a teleoperation 452 ought to be generated.
When one or more components of the AV system 22 (
In some implementations, a teleoperation event (422 in
A teleoperation event 422 generated by the AV system monitoring process 420 may comprise one or more of the following items of information:
The existence of circumstances suggesting the occurrence of an event need not be based on explicit information from the AV system 410 but can be inferred. For example, in some implementations, the AV system monitoring process 420 may determine or infer a failure in the AV system 410 by pattern recognition. For example, one or more signal values received from the AV system 410 that are out of a specified pattern may be determined as a system failure. Patterns can be hand-crafted or deduced from data via machine learning approaches such as re-enforcement learning or deep learning.
In some implementations, the AV system monitoring process 420 may detect a failure in the AV system 410 by a model-based approach. A model of the monitored hardware component or software process is constructed and a current state of the model is estimated using past inputs or past measurements. When a measurement associated with the current state deviates from its estimate, a system failure may occur. For example, dynamic quantities (e.g., velocity and orientation) of the AV with respect to throttle and steering commands is described in a dynamics model, and the monitoring process 420 uses the dynamics model to estimate the dynamic quantities at time t based on the throttle and steering commands at time t−1. When the measured dynamic quantities at time t differ from the estimated dynamic quantities by at least 1%, 2%, 3%, 4%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45% or 50%, the monitoring process 420 determines a system failure. A model may be hand-designed or identified using system identification approaches or learned using machine learning approaches (e.g., neural networks).
Teleoperation Event Handling Process.
Referring again to
The teleoperation event handling process 430 may generate a fallback request 432 and send it to the teleoperation command handling process 440. The fallback request 432 specifies one or more fallback operations for the AV system 410 to implement in response to the teleoperation events 422 while waiting for one or more teleoperations 452. Examples of fallback operations are described as follows.
In some implementations, a fallback request 432 may, for example, specify one or more of the following (and a wide variety of other actions and operations and combinations of them): keep traversing the current planned trajectory autonomously; change the goal to an AV service center and re-plan the trajectory to the new goal based on autonomous driving; follow autonomously the current trajectory with a slower velocity; re-plan a trajectory to stop at the closest location that is safe to stop; or autonomously decelerate until stopped.
Each fallback operation can have two main attributes: one or more required system processes (e.g., minimum required onboard processes) and a cost (e.g., a computed cost) of the fallback operation. Examples of system processes include maneuvering, data communications, database access, motion planning, perception, or sensing, or combinations of them. A cost represents how much the fallback operation deviates from a nominal autonomous driving mode. For example, an AV system without failure may drive at a nominal velocity (e.g., 40 mph); when a failure process occurs, a fallback request to keep traversing autonomously the current planned trajectory with a reduced velocity (e.g., 20 mph) may not need to invoke a motion planning process but may require at least perception and sensing processes so that the AV system can avoid hitting objects. The cost of this example may comprise how much the velocity is reduced from the nominal velocity of the AV system typically driving on the same road, and how much perception accuracy the AV system will sacrifice when executing the perception and sensing processes without invoking the motion planning process.
A cost of a fallback operation may be described, for example, as a function of the fallback operation, the teleoperation event, and the current operation status of the AV system. When a fallback request specifies two or more fallback operations, the costs of individual fallback operations are added, or weighted-summed. The selection of one or more appropriate fallback operations may be based on priority. Some implementations may utilize a decision tree to determine a hierarchy of the selection. In some implementations, the selection of one or more appropriate fallback operations to be included in the fallback request can be based on solving a combinatorial optimization problem. Some implementations of the selection may be based on a machine learning approach, where the best fallback operation or an optimal set of fallback operations is inferred from a database. The database may comprise past selections in various teleoperation events.
When receiving a teleoperation event, the teleoperation event handling process 430 may initialize a list of fallback operations from which to make its selection, and remove the fallback operations that cannot invoke required system processes or whose cost is beyond a threshold or both. When two or more fallback operations remain on the list, the one with the least cost may be selected. For example, a first fallback operation for which the AV system would traverse a new trajectory to a safe stopping place may require processes of sensing, perception, motion planning, and maneuvering to be functional. A second fallback operation for which the AV system immediately starts to slow down to a stop along an existing trajectory may require the maneuvering process to be operational. If all the required processes of the two fallback operations remain functional, their costs are compared to determine which fallback operation should be executed. If the motion planning process of the AV system is out of order, the second fallback operation would be chosen since the first fallback operation is infeasible.
The teleoperation event handling process 430 may send a teleoperation request 434 to the teleoperation server 420. When the teleoperation request 434 arrives at the teleoperation server 450, the server may place the teleoperation request 434 in a queue 451 to allocate an available human teleoperator 470. When the allocated teleoperator 470 becomes available, the teleoperation request 434 is presented on a teleoperation interface 460 to the teleoperator 470. Allocating teleoperators 470 to teleoperation requests 434 may be based on one or more of the following: time (e.g., peak or non-peak hours, seasons, day time, and night time), knowledge of or experience with the vehicle (e.g., vehicle make and model), or knowledge of or experience in the neighboring environment of the vehicle (e.g., country, state, city, town, street, and landmarks) and a language to be used (e.g., an oral communication may be used between a teleoperator and a user of the AV system; a sequence of texts may be presented to a user of the AV system).
The teleoperation request 434 may comprise one or more of the following: relevant information about an AV system failure or other condition, AV system information and data 412, the teleoperation event 422, important features, currently active teleoperation events, one or more teleoperations, and data of the AV system associated with each active teleoperation event.
The teleoperation event handling process 430 may initialize on a client or on the server 450, or both, a list of potential teleoperations. Each potential teleoperation is associated with one or more (e.g., required) hardware components or software processes or both. Potential teleoperations that have unmet requirements may be removed from the list. For example, on a teleoperation server 450, a teleoperator 470 may tele-interact with the AV system 410 through the teleoperation system and issue a teleoperation command 452 comprising a new trajectory, which may require the maneuver process and the perception process to be operational so that the AV system 410 can drive along the specified trajectory without hitting any object. The remaining potential teleoperations on the list may be ranked based on how easy they are for the teleoperator 470 to tele-interact with the AV system 410 with respect to current active teleoperation events. A tele-interaction able to address more active teleoperation events is ranked higher.
The teleoperator 470 may review the information on the interface 460 and issue one or more teleoperation commands 452. A teleoperation command 452 may be expressed at one or more levels. For example, a high-level command may be expressed in a spoken natural language, or a written natural language, or both, for example “turn right, go straight, and make a u-turn”. A middle-level command may be expressed as an alphanumeric string, for example, “a001, b005, a003”, where a001 is a code representing turning right, b005 representing going straight, and a003 representing making a u-turn. A low-level command may be expressed as machine instructions, for example,
Regardless of the level, the teleoperation command 452 may comprise a description of a behavior of the AV system 410, or one or more steps to be executed by the AV system 410, or both. When the teleoperation command handling process 440 receives the teleoperation command 452, it converts it into AV system commands 442 for controlling and maneuvering the AV system.
An AV system command 442 in general comprises machine instructions, for example, expressed in an assembly language or a low-level language, e.g., C/C++. When a teleoperation command 452 is expressed in a high-level language, such as a natural language, the teleoperation command handling process 440 may convert the teleoperation command 452 into machine instructions for the AV system 410.
Teleoperation Command Handling Process.
The teleoperation command handling process 440 handles fallback requests from the teleoperation event handling process 430 based on one or more teleoperation events 422, teleoperation commands 452 issued by the teleoperator 470 via the teleoperation interface 460, or both. In some implementations, a difference (e.g., a conflict) may exist between a fallback request 432 and a teleoperation command 452. For example, a fallback request 432 may ask the AV system 410 to operate at a reduced velocity along an existing trajectory, but simultaneously the teleoperation command 452 may ask the AV system 410 to operate at a nominal speed along a new trajectory. Thus, the teleoperation command handling process 440 has to mediate the difference to make sure the AV system 410 drives safely during a transition between a fallback operation and a teleoperation.
In some implementations, the teleoperator 470 may initiate a tele-interaction without a teleoperation request 434 having been generated. The teleoperator 470 may independently initiate a teleoperation command 452 to the teleoperation command handling process 440. For example, a weather condition may change from sunny to snowy, and the teleoperator may request the AV system 410 to drive back to an AV service center although the AV system monitoring process 420 has not generated any teleoperation event 422 in response to the weather change.
The teleoperation command handling process 440 takes a teleoperation command 452 issued by a teleoperator 470 through a teleoperation interface 460 and translates the teleoperation command 452 into one or more AV system commands 442. The AV system commands 442 are then sent to corresponding hardware components or software processes of the AV system 410.
In
When a teleoperation server 450 receives a teleoperation request 434, the teleoperation server 450 analyzes the teleoperation request 434 and the associated data, such as relevant information of a system failure, system information and data 412, the teleoperation event 422, important features, currently active teleoperation events, one or more teleoperations, or data of the AV systems associated with each active teleoperation event, or combinations of them. The teleoperation server 450 may present corresponding information to the teleoperator 470.
When a teleoperation request arrives at the communication interface 526 of the teleoperation server, the teleoperation request may be handled by a queuing process 532. In some implementations, the queuing process 532 may consider a first-in first-out method. In some cases, the queuing process 532 may evaluate the urgency of the teleoperation request, and then prioritize the urgent teleoperation request. A degree of urgency may be associated with safety. For example, an event that an AV system is under a fire may be placed with a high degree of urgency; a flat tire occurrence where the AV system has been parked in a safe place may be placed with a low degree of urgency.
Prioritizing a teleoperation request may utilize a decision tree to determine a hierarchy of existing teleoperation requests. In some implementations, prioritization can be based on solving a combinatorial optimization problem. Some implementations of the prioritization may be based on a machine learning approach analyzing a database; the database may comprise past teleoperation requests.
The teleoperation server 501 may comprise an interface manager 534, which renders content for a teleoperator to conduct a tele-interaction session. The teleoperator may conduct the tele-interaction on trajectory planning, where one or more trajectory primitives are used based on a primitive adjusting process 536 (whose details will be described below). When the teleoperator reviews relevant information, he may issue a teleoperation command. The teleoperation server may comprise a teleoperation command issuer 538 to communicate the command to the teleoperation command handling process of a teleoperation client. In some implementations, the teleoperation command issuer 538 may convert the teleoperation command into suitable machine instructions, e.g., alphanumeric strings or computer code.
A tele-interaction between a teleoperator and an AV system may rely on an interface apparatus. For example,
For example,
Referring to
Tele-Interaction with the AV System.
The teleoperation server may enable the teleoperator to interact with a hardware component or a software process of the AV system, for example, one or more of the autonomous driving capabilities. Different types of tele-interactions are allowed. For example, a tele-interaction on localization helps the AV system to identify the AV system's location when an onboard localization process fails; a tele-interaction on trajectory helps the AV system to identify a new trajectory or update an existing trajectory; a tele-interaction on annotation helps the AV system to recognize a perceived object. Many other examples exist.
Tele-Interaction on Localization.
When a localization component (i.e., a process that determines the geolocation of the AV) on the AV system fails, a teleoperation event for the failed localization is generated. The teleoperator may invoke tele-interaction with respect to localization for the AV system, which guides the AV system to re-localize itself. For example,
The information identifying the position of the spot 852 is transmitted within a teleoperation command back to the AV system. In some implementations, the spot 852 identified by the teleoperator may be treated by the teleoperation command handling process as a deterministic command. Thus, a motion planning process may resume with the spot 852 considered as a starting position and search for an optimal trajectory toward the original goal.
In some implementations, the spot 852 may be treated as a non-deterministic location, and the teleoperation command handling process may use probabilistic reasoning to identify a true geolocation on the map data. For instance, the spot 852 may be considered as prior knowledge, and a conditional probability on the prior knowledge may be computed to infer a true geolocation of the AV system. In some cases, the conditional probability may consider other information comprising one or more of the following: past or current or both perception data, past or current or both trajectory data, map data, sensing data from an onboard sensor, sensing data from an off-board sensor, and data from an external data source.
Tele-Interaction on Motion Planning.
When a motion planning process on the AV system fails, a teleoperation event for the failed motion planning process may be generated. The teleoperator may invoke tele-interaction for motion planning for the AV system, which guides the AV system to identify a trajectory.
For example,
The interface 910 may display a map around the AV 930 and a goal 932. The teleoperator may review the associated data and determine (e.g., draw) a new trajectory for the AV system 930 on the map. The interface 910 may switch to another interface 950 during the tele-interaction session and show a new trajectory 952 on the map data. A teleoperation command may comprise the new trajectory and may be sent to the teleoperation command handling process on the AV system.
In some implementations, the teleoperator provides one or more seeds 920 of a possible trajectory, and a new trajectory 952 is generated on the interface 950. A seed may be a point or a trajectory segment. A teleoperation command may comprise the one or more seeds, the new trajectory, or both, and be sent to the teleoperation command handling process on the AV system.
In a tele-interaction session, the teleoperator may interact with the motion planning process of the AV system. The teleoperator may perform one or more of the following:
In some implementations, a tele-interaction may specify one or more of the following elements. The specification may be determined by the teleoperator or computationally derived or both.
An interface for a tele-interaction on trajectory may rely on trajectory primitives for a teleoperator to generate or manipulate a trajectory. Referring to
A primitive may have a set of parameters that can be adjusted by the teleoperator. Examples of parameters may include one or more of the following: a segment length, a velocity of the AV system when entering the primitive, a velocity of the AV system driving along the primitive, a velocity of the AV when reaching the end of the primitive, allowance or prohibition of a lane change, a radius of a turn (e.g., left turn, right turn, and U-turn), a difference between a position (including orientation) at the beginning and the end of a turn, a maximum allowable yaw rotation rate of the AV during the traversal of the primitive, and an ending position of the primitive.
Referring
In some implementations, after a first primitive is selected and set by a teleoperator, the primitive adjusting process 536 may recommend options of feasible primitives that may be connected with the first primitive. When a second primitive is determined to be connected with the first primitive, the default parameter values of the second primitive may be automatically inferred by the primitive adjusting process 536 to ensure the compatibility (e.g., velocity, position and turn) across the connected primitives.
The primitive adjusting process 536 may utilize other data sources, such as map data, to appropriately set default values of the parameters. For example, an entry or exit velocity of a primitive may be set according to a speed limit of the road on which the AV system is; a default lateral offset of a lane-change maneuver may be set automatically according to the width of a lane where the AV is currently driving.
Referring
In some implementations, the teleoperation command handling process 440 may infer missing information. For example, a pair of positions (including orientations) at two locations may have been designated by the teleoperation command 452, but the connecting trajectory from one position to the other may be missing in the teleoperation command. The teleoperation command handling process 440 may, by itself or by invoking the motion planning process, generate a feasible connecting trajectory from one position to the other. Inferring the missing trajectory may be performed using a rule-based system that, for example, transforms a positional difference between the two positions into a smooth trajectory. Inferring the missing trajectory may be cast as an optimization problem in which variables are intermediate positions between the given pair of positions, and a cost function can be defined as positional differences between the intermediate positions; e.g., the cost function may be a sum of squares of positional differences. Minimizing the cost function will result in an optimal trajectory, which will ensure the resulting transition exhibits smooth and gradual changes in driving orientations.
In some implementations, a teleoperation command 452 may comprise a trajectory without a speed profile, the teleoperation command handling process 440 may, by itself or by invoking the motion planning process, generate a speed profile that leads to safe traversal of the trajectory by considering data from other data sources, such as positions and velocities of other objects (e.g., vehicles and pedestrians) from the perception processes and road information from the map. A speed profile may be derived by dynamic programing where velocity constraints are propagated backward from the end to the beginning of the trajectory according to safety and comfort constraints.
Tele-Interaction on Hardware Components or Software Processes.
When a teleoperation request arrives at a teleoperation server, the teleoperator may invoke tele-interaction on hardware components or software processes (e.g., autonomous driving capabilities) of the AV system. For example,
In some implementations, the interface 1200 may allow the teleoperator to zoom into a software process for editing one or more internal steps, or zoom into a hardware component for editing one or more subcomponents. For instance, the teleoperator may select the perception process 1204, and internal steps (e.g., segmentation 1222, object detection 1224, and object recognition and classification 1226) may be displayed. The teleoperator may select a step to view, create, change, edit, delete, enable, disable, invoke, or neglect a parameter or an algorithm of the step.
In some implementations, the interface 1200 may display sensors (e.g., LIDAR 1232 or vision sensor 1234) of the AV system. In some cases, the interface 1200 may allow the teleoperator to view, edit, enable or disable functionalities and parameters of the sensors. In some cases, the interface 1200 may allow the teleoperator to view, create, change, edit, delete, enable, disable, invoke, or neglect data acquired from the sensors.
Although the descriptions in this document have described implementations in which the teleoperator is a person, teleoperator functions can be performed partially or fully automatically.
Other implementations are also within the scope of the claims.
This application is a continuation of and claims priority to U.S. patent application Ser. No. 15/624,839, filed Jun. 16, 2017, the entire contents of which are incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 15624839 | Jun 2017 | US |
| Child | 16853713 | US |
