Method and system for dynamically curating autonomous vehicle policies

The system dynamically curates policies for autonomous vehicles by processing inputs and reducing the number of policies considered, addressing high computing demands and latency issues, thereby enhancing efficiency and safety.

JP7587604B2Active Publication Date: 2024-11-20MAY MOBILITY INC
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2022580347
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-07-01
Filing Date
2021-07-01
Publication Date
2024-11-20
Estimated Expiration
2041-07-01

AI Technical Summary

Technical Problem

The challenge in autonomous vehicles is the vast number of potential policies, which leads to high computing resource demands, increased latency, and inefficiencies in selecting the optimal policy due to the inability to simulate all options.

Method used

A system and method for dynamically curating policies by collecting and processing inputs to determine a subset of available policies, leveraging infrastructure devices, teleoperator input, and reducing the number of policies considered, thereby minimizing computing resources and latency.

Benefits of technology

This approach reduces computing resources and latency in policy selection, enhances policy execution efficiency, and improves safety and performance by prioritizing relevant policies based on real-time environmental conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007587604000001
    Figure 0007587604000001
  • Figure 0007587604000002
    Figure 0007587604000002
  • Figure 0007587604000003
    Figure 0007587604000003
Patent Text Reader

Abstract

A system for dynamic policy curation that includes a computing system and that interfaces with an autonomous agent. The method for dynamic policy curation includes collecting a set of inputs; processing the set of inputs; and determining a set of available policies based on processing the set of inputs. Additionally or alternatively, the method may include any or all of selecting a policy; implementing the policy; and / or any other suitable process.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 046,842, filed July 1, 2020, the entirety of which is incorporated by reference herein.

[0002]

[0002] The present invention relates generally to the field of autonomous vehicles, and more specifically to novel and useful systems and methods for dynamically curating behavior policies for autonomous agents in the field of autonomous vehicles. [Background technology]

[0003]

[0003] State-of-the-art vehicle automation enables vehicles to operate effectively and safely in autonomous and / or semi-autonomous states. The inventors have made this possible through a multi-policy decision-making process that allows multiple candidate policies for an autonomous agent at a future time to be simulated, whereby an optimal policy is selected based on the simulation results.

[0004]

[0004] However, a technical problem that arises is that there may be a wealth of potential policies that a vehicle can choose from, which requires extensive computing resources to simulate and ultimately select an optimal policy (e.g., for safety, rider enjoyment, efficiently reaching a destination, etc.). These computing requirements may preclude simulating all candidate policies, which would require much more extensive computing resources, increase the latency between policy selection and execution, and / or cause other effects.

[0005]

[0005] The present inventors have discovered a novel and useful system and method, described below, for dynamically curating these policies. [Brief description of the drawings]

[0006] [Figure 1]

[0006] FIG. 1 is a schematic diagram of a dynamic policy curation system. [Diagram 2]

[0007] FIG. 1 is a schematic diagram of a dynamic policy curation method. [Diagram 3]

[0008] 1 illustrates a variation of a system and method for implementing infrastructure devices defining detection range sensing regions, where information is collected by the infrastructure devices and used in determining available policies. [Figure 4]

[0009] 1 shows a set of example policies available to an autonomous agent that are simulated in a multi-policy decision-making process before selecting an optional policy. [Figure 5A]

[0010] 1 illustrates an example map with various area options defined within the map, with each area associated with a particular set of available policies and / or rules associated with determining the available policies. [Figure 5B]

[0010] An example map is shown having various area options defined within the map, with each area associated with a particular set of available policies and / or rules associated with determining the available policies. [Figure 6A]

[0011] 13 shows a schematic variation of dynamic policy curation based on scenario decisions and / or teleoperator input. [Figure 6B]

[0011] FIG. 1 shows a schematic variant of dynamic policy curation based on scenario decisions and / or teleoperator input. [Figure 6C]

[0011] FIG. 1 shows a schematic variant of dynamic policy curation based on scenario decisions and / or teleoperator input. [Figure 7]

[0012] 1 shows a schematic variation of a set of inputs and outputs involved in a dynamic policy curation method; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0007]

[0013] The following description of preferred embodiments of the invention is not intended to limit the invention to those preferred embodiments, but rather to enable any person skilled in the art to make and use the invention.

[0008] 1. Overview

[0014] As shown in FIG. 1, the dynamic policy curation system 100 includes a computing system and interfaces with an autonomous agent. The system 100 may further include and / or interface with any or all of: a set of infrastructure devices, a communication interface, a teleoperator platform, a sensor system, a positioning system, a guidance system, and / or any other suitable components. Additionally or alternatively, the system may include all of the systems, components, embodiments, and examples described in U.S. Application No. 16 / 514,624, filed July 17, 2019, U.S. Application No. 16 / 505,372, filed July 8, 2019, U.S. Application No. 16 / 540,836, filed August 14, 2019, and U.S. Application No. 16 / 792,780, filed February 17, 2020, each of which is incorporated by reference in its entirety.

[0009]

[0015] 2, dynamic policy curation method 200 includes collecting S210 a set of inputs; processing S220 the set of inputs; and determining S230 a set of available policies based on processing the set of inputs. Additionally or alternatively, method 200 may include any or all of the following steps: determining a risk associated with one or more policies; determining that there are too many policies available and reducing the number of available policies; implementing S240 the policies; and / or any other suitable process.

[0010]

[0016] Additionally or alternatively, the methods may include all of the methods, processes, embodiments and examples described in U.S. Application No. 16 / 514,624, filed July 17, 2019, U.S. Application No. 16 / 505,372, filed July 8, 2019, U.S. Application No. 16 / 540,836, filed August 14, 2019, and U.S. Application No. 16 / 792,780, filed February 17, 2020, each of which is incorporated by reference in its entirety.

[0011] 2. Advantages

[0017] Systems and methods for dynamically curating autonomous vehicle policies can provide several advantages over current systems and methods.

[0012]

[0018] In a first variant, the system and / or method provides the advantage of reducing or minimizing the computing resources required to select a policy (also referred to herein as an equivalent behavior policy) for an autonomous vehicle. This may further enable any or all of: reducing latency in policy selection and / or execution; minimizing the occurrence of failed or less successful policy implementations; and / or any suitable functionality. In a specific example, the number of policies available to the autonomous agent is dynamically and intelligently adjusted based on any number of conditions, which serve to prioritize the most important policies for the agent to consider and eliminate lower priority policies or policies predicted to have a low success rate.

[0013]

[0019] In a second variation, additionally or alternatively to the first variation, the system and / or method provides the advantage of making many policies generally available to the autonomous agent, while prioritizing the subset to consider in each selection cycle. In a specific example, the set of policies available to the autonomous agent is large and is managed by dynamically curating the subset available to the vehicle based on any number of conditions and parameters.

[0014]

[0020] In a third variation, additionally or alternatively to the variations described above, the system and / or method may advantageously utilize input from a remote human operator (e.g., a teleoperator) in certain instances of dynamic policy curation. In particular instances, the teleoperator input may be used to any or all of: add and / or approve policies for consideration by the agent; remove policies for consideration by the agent; override policies and / or the policy selection process in the selection step; and / or enable any other suitable action.

[0015]

[0021] In a fourth variation, additionally or alternatively to the variations described above, the system and / or method provides the advantage of leveraging information from a set of external sensing devices (also referred to herein as equivalently roadside units and / or infrastructure devices) arranged in the autonomous agent's environment that can inform the agent's selection of available policies.

[0016]

[0022] Additionally or alternatively, the systems and methods may provide any other advantages.

[0017] 3. System 100

[0023] System 100 functions to curate and select policies available to an autonomous agent as the vehicle's environment dynamically changes. Additionally or alternatively, system 100 may function to select an optimal policy for the autonomous agent, execute the policy, and / or perform any other suitable function.

[0018]

[0024] System 100 includes a computing system and interfaces with an autonomous agent. System 100 may further include and / or interface with any or all of a number of infrastructure devices, communication interfaces, teleoperator platforms, and / or any other suitable components.

[0019]

[0025] The autonomous agent, also referred to herein equally as an autonomous vehicle (e.g., unmanned automobile, unmanned shuttle, unmanned automobile-grade vehicle, etc.) and / or ego-agent, is preferably a fully autonomous vehicle, but may additionally or alternatively be any semi-autonomous or fully autonomous vehicle (e.g., boat, unmanned aerial vehicle, unmanned automobile, etc.). Additionally or alternatively, the autonomous agent may be a vehicle that switches between semi-autonomous and fully autonomous states (or fully attended states), and thus the autonomous agent may have attributes of both semi-autonomous and fully autonomous vehicles depending on the state of the autonomous agent. Additionally or alternatively to an autonomous vehicle, the autonomous agent may include any type of autonomous machine, autonomous device, autonomous robot, and / or any other suitable device.

[0020]

[0026] The computing system is operable to collect information related to dynamic curation of policies for the vehicle. Additionally or alternatively, the computing system may be operable to select policies to be implemented, execute the policies, and / or perform any other suitable functionality.

[0021]

[0027] A policy preferably refers to behaviors and / or actions and / or movements associated with an agent. In a specific example, for example, a policy may refer to a particular behavior of an agent, such as, but not limited to, lane changing, lane keeping (e.g., going straight), merging, parking, turning (e.g., turning right, turning left, etc.), pulling onto a shoulder, stopping, starting, braking, accelerating, and / or any other behavior. Additionally or alternatively, a policy may refer to behaviors and / or actions and / or movements associated with a particular set of parameters, such as a behavior having a particular speed or set of speeds, a particular distance to a vehicle ahead, a particular braking distance, a particular acceleration and / or deceleration, and / or any other parameter. In a specific example, for example, a policy may refer to behaviors associated with a particular set of parameters associated with a behavior (e.g., going straight at a predefined speed, braking with a predefined braking distance, braking with a predefined deceleration, accelerating with a predefined acceleration and / or speed, etc.). Additionally or alternatively, a policy may include and / or refer to any other information.

[0022]

[0028] Additionally or alternatively, any other objects in the own agent's environment (e.g., other vehicles, pedestrians, animals, cyclists, dynamic objects, static objects, etc.), which are equivalently referred to herein as monitored objects, may be associated with (e.g., assigned to) one or more policies. In a preferred variant, these policies are used in the MPDM module (e.g., as described below) along with policies for the own agent in, e.g., estimating the intention of these objects in a set of simulations. Additionally or alternatively, the policies of the monitored objects may be used in other ways.

[0023]

[0029] The computing system preferably includes an on-board computing system located onboard the autonomous agent (e.g., integrated within the autonomous agent). Additionally or alternatively, the computing system may include any or all of: a remote computing system (e.g., a cloud computing system in lieu of an on-board computing system, a remote computing system in communication with the on-board computing system, etc.), a computing system integrated into an auxiliary device (e.g., a mobile device, a user device, etc.), an edge device including a mobile computing device, and / or any other suitable computing system and device. In one variation, for example, the autonomous agent is operable to communicate with a remote or distributed computing system, which may include a user device (e.g., a mobile phone, a laptop, etc.), a remote server, a cloud server, or any other suitable local and / or distributed computing system remote from the vehicle. The remote computing system may be connected to one or more systems of the autonomous agent through one or more data connections (e.g., channels), but may alternatively communicate with the vehicle systems in any suitable manner.

[0024]

[0030] The computing system preferably includes a processing system (e.g., a graphics processing unit or GPU, a central processing unit or CPU, or any suitable processing circuitry) and a memory, but may additionally or alternatively include any other suitable components. The memory may be short-term memory (e.g., volatile, non-volatile, random access memory or RAM, etc.) and / or long-term memory (e.g., flash memory, hard disk, etc.). In some variations, for example, the on-board computing system operates to interact with and / or operably control any or one or more of the identified components or modules described herein. For example, the on-board computing system may function to implement and / or execute computer instructions for implementing a multi-policy decision module, a synchronization module, etc. In a specific example, the processing system and memory function collectively to dynamically manage a set of policies (e.g., determined based on method 200) available to an autonomous agent within the framework of a multi-policy decision-making framework, such as those described in U.S. Application No. 16 / 514,624, filed July 17, 2019, which is incorporated herein by reference in its entirety. Additionally or alternatively, the processing system and memory and / or any other suitable components may be used for any other suitable functionality.

[0025]

[0031] A computing system (e.g., an on-board computing system) preferably functions to control the autonomous agent and to process detected data from a sensor suite (e.g., computer vision system, LIDAR, flash LIDAR, wheel speed sensors, GPS, etc.) of the autonomous agent and / or other sensors (e.g., from infrastructure devices) to determine the state of the autonomous agent and / or the state of the agents in the autonomous agent's operating environment. Based on the state and program instructions of the autonomous agent and / or the agents in the operating environment, the on-board computing system preferably modifies or controls the behavior of the autonomous agent, such as through selection of a policy. Additionally or alternatively, the computing system may include a multi-policy decision-making module that functions to generate policies and select policies (e.g., change lanes, merge, stay in current lane, turn left, turn right, stop, slow down, accelerate, stop at a traffic light, stop at a stop sign, yield, etc.) that the computing system can execute to control the behavior of the autonomous agent.

[0026]

[0032] In a first set of variations, the computing system includes an on-board general-purpose computer adapted for I / O communication with the vehicle control system and the sensor system, but may additionally or alternatively be any suitable computing device. The on-board computing system is preferably connected to the Internet via a wireless connection (e.g., via a cellular link or connection). Additionally or alternatively, the on-board computing system may be coupled to any number of wireless or wired communication systems.

[0027]

[0033] System 100 preferably includes a communications interface for communicating with the computing system, the communications interface functioning to allow information to be received (e.g., from an infrastructure device, from a remote computing system and / or remote server, from a teleoperator platform, another autonomous agent or other vehicle, etc.) and to transmit information to the computing system (e.g., to a remote computing system and / or remote server, to a teleoperator platform, to an infrastructure device, another autonomous agent or other vehicle, etc.). The communications interface preferably includes a wireless communication system (e.g., Wi-Fi, Bluetooth, Cellular 3G, Cellular 4G, Cellular 5G, Multiple Input Multiple Output or MIMO, one or more radios, or any other suitable wireless communication system or protocol), but may additionally or alternatively include any or all of: a wired communication system (e.g., modulated power line data transfer, Ethernet, or any other suitable wired data communication system or protocol), a data transfer bus (e.g., CAN, FlexRay), and / or any other suitable components.

[0028]

[0034] The system 100 may include a set of infrastructure devices operable to observe one or more aspects and / or characteristics of the environment and to collect observational data relating to one or more aspects and / or characteristics of the environment. In a preferred variant, the infrastructure devices additionally operate to collect data associated with the observations and to transmit the collected data and / or processed derivatives of the collected data to the autonomous agent. Additionally or alternatively, the infrastructure devices may operate to collect and transmit data to a teleoperator platform, which may use the data to inform teleoperator decision-making, such as whether to include and / or exclude policies from consideration by the autonomous agent's computing system. In a specific example, the infrastructure devices may, for example, enable visibility around corners of the vehicle that the teleoperator may use to enable turn policies to be considered by the autonomous agent (by verifying that the road is clear for the turn).

[0029]

[0035] In a first set of variations, for example, the infrastructure device forwards collected observation data to an autonomous vehicle service and / or a remote platform (e.g., implemented via a network of distributed computing systems) that operates to interactively communicate with the autonomous agent and / or control one or more functions of the autonomous agent.

[0030]

[0036] The infrastructure devices are equivalently referred to herein as roadside units, which preferably include devices within the immediate vicinity and / or proximity or short-range communication proximity of the autonomous agent's operating location and are operable to collect data regarding conditions within the area surrounding the autonomous agent and proximate to the autonomous agent's operating zone. In an embodiment, the roadside units include one or more of off-board detection devices including flash LIDAR, thermal imaging devices (thermal cameras), still or video capture devices (e.g., image cameras and / or video cameras, etc.), global positioning systems, radar systems, microwave systems, inertial measurement units (IMUs), and / or any other suitable detection device or combination of detection devices.

[0031]

[0037] The infrastructure device may include computing capabilities via processing circuitry and communications interfaces that enable the infrastructure device to communicate with any or all of an autonomous agent's computing system, a remote computing system, a teleoperator platform, and / or any other suitable component or combination of components.

[0032]

[0038] The zone of operation of the autonomous agent may optionally be defined as a predefined radius (e.g., 100 feet, 50 feet to 100 feet, less than 50 feet, 100 feet to 200 feet, more than 200 feet, etc.) along the structured and / or unstructured path of the autonomous agent at any point along the structured path where the autonomous agent 110 is positioned and / or operating (e.g., traveling). In a specific example of a structured and / or predefined autonomous agent path, the proximity zone of operation of the autonomous agent is 100 feet from or along any point along the structured path.

[0033]

[0039] Technical advantages achieved by the implementation of infrastructure devices may include the ability to observe situations beyond the autonomous agent's observable range (e.g., around a corner, a right-angled road going down, etc.), which the autonomous agent can subsequently act in curating one or more policies available to the agent. In certain instances, for example, observations of one or more aspects of a given environment may be made by the autonomous agent, and observations of one or more different and / or overlapping aspects of the given environment may be made from different perspectives by one or more infrastructure devices disposed and operating in the given environment. In such embodiments, the infrastructure device's perspective including observation data therefrom is augmented with observation data from the autonomous agent's perspective to generate a comprehensive view of the autonomous agent's operating environment and / or to provide additional perspectives to one or more teleoperators of a teleoperator platform. This allows for improved predictions of the operating environment and improved policy decisions selected and / or executed by the autonomous agent operating independently and safely (from an on-board human operator) within the operating environment.

[0034]

[0040] In some variations, the autonomous agent may augment and / or fuse data derived by an on-board sensor suite (e.g., as described below) with additional observations from infrastructure devices (e.g., roadside units) to improve policy curation and / or selection by the autonomous agent.

[0035]

[0041] Additionally or alternatively, the infrastructure device may detect and track any type or kind of agent in the operating environment, such as with a video camera or radar. In one variation, for example, the video camera may function to detect the agent and provide a semantic classification of the agent type and possible intent of the agent, such as a pedestrian about to cross the road or a car about to turn left, a driver about to open a car door and exit a vehicle, a cyclist traveling in a bike lane, and / or any other suitable information.

[0036]

[0042] Additionally or alternatively, any or all of the infrastructure devices may include traffic management devices (e.g., traffic sensors, traffic signals, pedestrian signals, etc.) operating in the environment that may: communicate with any or all of the other infrastructure devices (e.g., roadside units); may function to communicate directly with an autonomous agent; with a teleoperator platform; and / or in any other suitable manner; regarding any or all of the data collected and / or detected by the infrastructure devices, regarding the operational status of the infrastructure devices (e.g., red or green light) and / or any other information. In a specific example, a traffic light may be an infrastructure device in the environment surrounding the autonomous vehicle that may function to communicate directly with the autonomous vehicle or a roadside unit that may be in operative communication with the autonomous vehicle. In this example, the traffic light may function to share and / or communicate operational status information, such as the color of light the traffic light is projecting, or other information, such as the timing of light changes by the traffic light.

[0037]

[0043] In a preferred variant, the communication interface allows the autonomous agent to communicate and / or exchange data with systems, networks and / or devices external to the autonomous agent. This communication interface and / or a separate communication interface further preferably allows one or more infrastructure devices to communicate directly with the autonomous agent and / or remote computing systems and / or teleoperator platforms. The communication interface preferably includes one or more of a cellular system (or any suitable long-range communication system), direct short-wave radio, or any other suitable short-range communication system.

[0038]

[0044] System 100 preferably includes a sensor suite including sensors onboard the autonomous vehicle (e.g., computer vision systems, LIDAR, RADAR, wheel speed sensors, GPS, cameras, etc.) that communicate with the onboard computing system and function to gather information for dynamically curating policies available to the autonomous agent. Additionally or alternatively, the sensor suite may function to enable autonomous agent operation (e.g., self-driving), collect data regarding the situation surrounding the autonomous agent, collect data regarding the operation of the autonomous agent, detect the need for maintenance of the autonomous agent (e.g., via engine diagnostic sensors, external pressure sensor strips, sensor health sensors, etc.), detect cleanliness standards within the autonomous agent (e.g., internal cameras, ammonia sensors, methane sensors, alcohol vapor sensors), and / or perform any other suitable function.

[0039]

[0045] The system may optionally include one or more controllers and / or control systems operable to control the autonomous movement and / or actions of the autonomous agent based on commands generated by the one or more computing systems. Additionally or alternatively, the one or more controllers may be operable to implement machine learning techniques to assist the controller's functions, such as feature detection / classification, obstacle mitigation, path traversal, mapping, sensor integration, ground truth determination, and / or enable any other suitable functionality. The controller may include any suitable software and / or hardware components (e.g., a processor and computer-readable storage) utilized to generate control signals for controlling the autonomous agent in accordance with the autonomous agent's routing goals and the autonomous agent's selected policies.

[0040]

[0046] In one variation, the control system includes a dedicated arbiter of control signals that selectively passes one or more control signals to one or more lower level driving mechanisms of the autonomous agent, such as any of the dedicated arbiters (e.g., arbiter circuits) described in U.S. Application No. 16 / 540,836, filed August 14, 2019, which is incorporated herein by reference in its entirety. The arbiter may additionally or alternatively function to manage and / or implement policies, such as through methods described below.

[0041]

[0047] Additionally or alternatively, the system may include any or all of a sensor fusion system, a positioning system (e.g., including position sensors of a sensor system), a guidance system, and / or any suitable components. In one variation, for example, the sensor fusion system synthesizes and processes sensor data and, together with a multi-policy decision module, predicts the presence, location, classification, and / or path of objects and characteristics of the autonomous agent's environment. In various embodiments, the sensor fusion system may function to incorporate data from multiple sensors and / or data sources, including, but not limited to, cameras, LIDAR, radar, infrastructure devices, remote data feeds (Internet-based data feeds), and / or any number of other types of sensors.

[0042]

[0048] The positioning system processes the sensor data along with other data to determine the location of the autonomous agent relative to the environment (e.g., local position relative to a map, precise position relative to road lanes, vehicle heading, speed, etc.), which may serve to determine what policies are available to the autonomous agent (e.g., as described below). The guidance system processes the sensor data along with other data to determine the path the vehicle should follow.

[0043]

[0049] The system may optionally interface with a teleoperator platform that references one or more remote teleoperators and associated components (e.g., communication interfaces with autonomous agents, computing systems, output devices for displaying information from the autonomous agents and / or infrastructure devices to the teleoperator, input devices for receiving instructions / commands from the teleoperator, etc.) The teleoperator platform may function to receive inputs from the teleoperator, which may be used at least in part in determining the curated policy for the vehicle.

[0044]

[0050] Additionally or alternatively, system 100 may include any other suitable components.

[0045] 4. Method 200

[0051] 2, dynamic policy curation method 200 includes collecting S210 a set of inputs; processing S220 the set of inputs; and determining S230 a set of available policies based on processing the set of inputs. Additionally or alternatively, method 200 may include any or all of the following steps: determining a risk associated with one or more policies; determining that there are too many policies available and reducing the number of available policies; implementing S240 the policies; and / or any other suitable process.

[0046]

[0052] Method 200 functions to dynamically determine which policies are available to an autonomous agent, which may subsequently enable any or all of: reducing the computing resources required to select a policy; ensuring that the policy selected by the vehicle is optimal for the autonomous agent's environment; availability of a diverse and numerous set of generic policies available to the autonomous agent; and / or performance of any other suitable function.

[0047]

[0053] Method 200 is preferably performed continuously throughout the operation of the autonomous agent, such as every selection step (e.g., every decision-making step) of the autonomous agent. Additionally or alternatively, method 200 may be performed with a predefined frequency, in response to a trigger (e.g., detection of a change in the autonomous agent's environment based on teleoperator input, detection of a transition in the autonomous agent's map area, detection of a change in the autonomous agent's operating conditions, etc.); at random time intervals; and / or at any other suitable time.

[0048]

[0054] In specific examples, method 200 is performed according to a selection cycle associated with a predefined frequency (e.g., 4 times per second, 1 time per second, 2 times per second, 3 times per second, 5 times per second, 1-5 times per second, 1-10 times per second, 10-50 times per second, more than 50 times per second, etc.). Additionally or alternatively, the selection cycle may be defined based on a random interval and / or may be defined in other manners. Still additionally or alternatively, method 200 may be performed in other manners.

[0049]

[0055] The method 200 is preferably performed by the system 100 as described above, but may additionally or alternatively be performed by any suitable system.

[0050]

[0056] Performing method 200 preferably includes implementing one or more rule-based processes, such as within the context of a multi-policy decision-making process, but may additionally or alternatively be implemented in other manners, such as through one or more machine learning processes, including but not limited to deep learning processes (e.g., neural networks, convolutional neural networks, etc.) and / or any other suitable processes.

[0051] 4.1 Method - Step S210 of Collecting a Set of Inputs

[0057] Method 200 includes a step S210 of collecting a set of inputs, which serves to appropriately determine and / or prioritize what policies are available to the autonomous agent. Additionally or alternatively, S210 may serve to select a policy, provide information to one or more entities (e.g., a teleoperator platform) for use in decision making, and / or perform any other suitable function.

[0052]

[0058] S210 is preferably performed initially in method 200, and optionally multiple times during operation of the autonomous agent, such as continuously, at a predefined frequency (e.g., at each selection cycle), at a set of predefined intervals, at a set of random intervals, and / or at any other time. Additionally or alternatively, S210 may be performed in response to a trigger, once during method 200, in response to another process of method 200, in parallel with another process of method 200, and / or at any other time.

[0053]

[0059] In a preferred set of variations, for example, the set of inputs is collected continuously throughout the vehicle movement (e.g., at each selection step of the own agent's selection cycle), but may additionally or alternatively be collected at any of the times mentioned above.

[0054]

[0060] The inputs are preferably collected from at least a sensor system of the autonomous agent (e.g., a vehicle position sensor, a set of vehicle mounted cameras, a set of vehicle mounted radar sensors, a set of vehicle mounted lidar sensors, etc.), but may additionally or alternatively be collected from any or all of a sensor system of an infrastructure device, a teleoperator platform, a third party sensor, one or more databases (e.g., a third party database, a weather database, a traffic database, etc.), and / or from any other suitable source.

[0055]

[0061] In a preferred variation, the input includes a vehicle position (e.g., continuously collected), which is used to locate the vehicle in one or more maps described below, and which is then used to dynamically curate the agent's policies based on the agent's position. The vehicle position is preferably in the form of position coordinates (e.g., GPS coordinates), which are then referenced using one or more maps (e.g., as described below), although any other suitable position information may additionally or alternatively be included. Additionally or alternatively, the vehicle position may be determined based on other sensor information (e.g., camera information, other information (e.g., vertical characteristics of the road surface)), and / or the vehicle position may be determined in any other suitable manner.

[0056]

[0062] Additionally or alternatively, any other suitable sensor information may be collected, such as any or all of: motion information (e.g., agent velocity and / or speed, agent acceleration); directional information (e.g., agent yaw angle, attitude, etc.); temporal information (e.g., from a clock, to determine whether the vehicle is traveling during day or night, to determine if the vehicle is traveling during rush hour, etc.); environmental information (e.g., from one or more cameras and / or radar sensors and / or lidar sensors); and / or any other suitable information.

[0057]

[0063] One or more inputs related to the agent's environmental conditions (e.g., lighting, rain, etc.) are further preferably collected (e.g., continuously), which defines the agent's situational awareness and can function to include policies for consideration, exclude policies from consideration, and / or curate and / or prioritize the set of policies available to the agent. In a specific example, for example, determining that the field of view of one or more sensors (e.g., optical sensors, cameras, radar sensors, lidar sensors, etc.) is obstructed based on the environmental conditions (e.g., because of glare, because of poor lighting conditions, because of heavy rain, because of snow, because of objects obstructing the field of view, etc.) can be used to determine whether some behavioral policies should be excluded from consideration (e.g., turn right when the right camera is obstructed), whether some policies should be added to consideration (e.g., policies to clear obstacles, change lanes to reduce glare, etc.), whether other inputs should be received (e.g., from a teleoperator), and / or these inputs can be used in other ways. These inputs (and any other inputs described herein) may be collected from any or all of: sensor systems; third party tools (e.g., client applications), sites and / or databases (e.g., online weather information, real-time traffic information, etc.); user inputs (e.g., aggregated user inputs, teleoperator inputs, on-board operator inputs, etc.); historical information; and / or information collected from any other suitable source.

[0058]

[0064] Additionally or alternatively, one or more inputs may be collected regarding the vehicle's surroundings, such as characteristics of the monitored object (e.g., as described above) (e.g., location, speed, predicted characteristics, etc.); the proximity of the monitored object to the own agent (e.g., surrounding vehicles, nearby pedestrians, etc.); potential obstacles to the own agent; the proximity of the own agent to road landmarks (e.g., traffic lights, lane boundaries, etc.); inputs collected from infrastructure devices; and / or any other suitable inputs.

[0059]

[0065] In one variation, inputs are collected that are used to determine, for example, one or more features associated with a monitored object (e.g., a vehicle, a pedestrian, other object, etc.) in the own agent's environment. The features preferably include a position and velocity associated with the monitored object, but may additionally or alternatively include any or all of an orientation (e.g., heading, attitude, etc.), acceleration, and / or any other features. The features may include any or all of an actual feature, an estimated feature, a predicted feature (e.g., a feature of a future monitored object, etc.), and / or any other features. The inputs associated with the monitored object are preferably collected from the own agent's sensor system (e.g., a camera, a radar sensor, a lidar sensor, a proximity sensor, etc.), but may additionally or alternatively be collected using another sensor system (e.g., mounted on an infrastructure device, mounted on the monitored object itself, etc.), a database, received from the monitored object itself, and / or otherwise received from any other suitable source.

[0060]

[0066] In a set of illustrative examples, features associated with the monitored objects are ultimately used by the MPDM module (e.g., as described below) to select the optimal policy for the agent (e.g., through intent inference of the monitored objects in a set of simulations run by the MPDM module).

[0061]

[0067] Additionally or alternatively, one or more inputs regarding the health of vehicle components may be collected, such as any or all of: information regarding sensor health (e.g., of the autonomous agent, of infrastructure devices, etc.) from one or more health monitors, such as those described in U.S. Application No. 16 / 505,372, filed July 8, 2019, which is incorporated by reference in its entirety; performance characteristics associated with components of the autonomous agent (e.g., braking system, steering system, control system, etc.) and / or infrastructure devices; and / or any other suitable information.

[0062]

[0068] In some variations (e.g., as described above), sensors of one or more infrastructure devices (e.g., roadside units) may be used to communicate any or all of the following information: knowledge of the state of a particular traffic light; knowledge of the agent state and the agent's surroundings or potential surroundings (e.g., field of view around corners that the agent cannot detect from its on-board sensors); and / or any other suitable information. As part of dynamically curating policies, in addition to dynamically curating policies and / or as an alternative to dynamically curating policies, the infrastructure device may function to evaluate risks associated with potential policies. This may function to allow a particular policy to remain in the set of curated policies, to remove a policy from the set of curated policies, to determine whether a teleoperator's input should be triggered, and / or to perform any other suitable function. In a first set of examples, the infrastructure device may function to add, retain, and / or remove one or more policies from the curated set for consideration by its own agent (e.g., in the MPDM module). In one example, a policy that may have been eliminated from the set because, for example, the view of the own agent's camera is occluded may remain for consideration if the infrastructure device's camera has a view of the occluded area (e.g., indicating that no surveillance object is present in the occluded area). In another example, a policy that may have been included in the set may be eliminated from consideration based on, for example, the infrastructure device detecting that there is heavy traffic or other delays (e.g., construction zone) in an area that the agent will encounter in implementing that policy (e.g., making a right turn). Additionally or alternatively, the infrastructure device may be used in other ways to signal any or all of the policies.

[0063]

[0069] The input may optionally include one or more commands or other information from an on-board operator, such as a safety driver, and / or a remote operator, such as a teleoperator. This may include permission for the agent to consider and / or select policies that the agent cannot or would not select itself, such as high-risk policies and / or otherwise rare policies. In a specific example, for example, the teleoperator provides an input that the agent may overtake a slowly moving vehicle in front of it if the agent determines it is safe to do so. The command and / or permission received from the operator is preferably processed with other available policies in subsequent processes of the method and / or in a multi-policy decision-making module of the system (e.g., simulated with other available policies), but may additionally or alternatively be selected without further processing (e.g., in an override manner).

[0064]

[0070] In one variation, an alert is sent to the teleoperator in response to detecting a particular scenario, using a classification module (eg, as described below), and input from the teleoperator is requested.

[0065]

[0071] The set of inputs may additionally or alternatively include any or all of: inputs from one or more computing systems (e.g., as described above), historical information (e.g., learned fleet knowledge, aggregated information, etc.), information from one or more servers and / or databases (e.g., lookup tables), and / or any other suitable inputs.

[0066]

[0072] In a first variant, S210 includes a step of receiving a set of sensor inputs from a sensor system mounted on the own agent, the set of sensor inputs including at least location information associated with the own agent, which location information may be used in a subsequent process of the method for determining a set of policies available to the agent (e.g., based on a map).

[0067]

[0073] The set of inputs may additionally or alternatively be received from any other sensor system (e.g., sensors of a set of infrastructure devices, sensors from other agents, etc.), an information source (e.g., a database, a third party system, etc.), a teleoperator, and / or any other information source.

[0068]

[0074] The set of inputs may additionally or alternatively include any or all of: motion information associated with the own agent, such as speed information, acceleration information, and / or any other motion information; directional information (e.g., heading, attitude, etc.); sensor health information; history information; and / or any other information.

[0069]

[0075] The set of inputs may additionally or alternatively include inputs for determining information associated with one or more monitored objects in the own agent's environment (e.g., location information, orientation information, motion information, etc.) The inputs associated with the monitored objects are preferably collected at least in part by a sensor system onboard the own agent, but may additionally or alternatively be received from infrastructure devices, sensors onboard the monitored objects, and / or any other components.

[0070]

[0076] In a specific example of the first variant, S210 includes a step of collecting a set of sensor inputs from a set of sensors of the own agent, including a set of cameras mounted and positioned on the own agent (e.g., mounted externally of the own agent, mounted internally of the own agent, mounted on both, etc.), and optionally any other sensors mounted on the agent (e.g., position sensors, speedometers, accelerometers, inertial measurement units, etc.), and the sensor inputs are used to determine a set of policies associated with the own agent in subsequent processes of the method.

[0071]

[0077] Additionally or alternatively, S210 may include any other suitable process.

[0072] 4.2 Method: Step S220 of Processing a Set of Inputs

[0078] Method 200 includes a step S220 of processing the set of inputs, which serves to determine which policies are available to the agent. Additionally or alternatively, S220 may serve to trigger the collection of further information (e.g., from a teleoperator) and / or perform any other function.

[0073]

[0079] S220 is preferably performed in response to and based on S210, but may additionally or alternatively be performed in response to, prior to, in parallel with, and / or at any other suitable time, any other process of method 200. Additionally or alternatively, S220 may be performed multiple times during operation of the own agent (e.g., in a predefined selection cycle, in response to each iteration of S210, continuously, at predefined frequencies, at random intervals, etc.), in response to a trigger, and / or at any other time.

[0074]

[0080] S220 is preferably performed using the agent's on-board computing system and / or processing system (e.g., as described above), but may additionally or alternatively be performed by other suitable computing systems (e.g., a remote computing system), a teleoperator or an on-board operator, and / or may be performed using any other suitable components.

[0075]

[0081] S220 may include processing the set of inputs received in S210 to determine a set of features and / or parameters associated with any or all of the own agent, the set of monitored objects, the set of environmental conditions, and / or any other features and / or parameters. This may include, for example, processing raw sensor data to determine (e.g., derive, calculate based on a set of equations and / or algorithms, determine using a set of models, etc.) any or all of the information described above (e.g., motion information, location information, etc.) and / or any other features or parameters. The inputs may be processed using any or all of a set of equations (e.g., dynamic equations), algorithms, rules, models, lookup tables and / or databases, and / or any other tools.

[0076]

[0082] S220 preferably includes processing the set of inputs (and / or the features or parameters determined based on the set of inputs) with a set of rules, which set of rules specifies which policies the agent can use and / or which policies the agent cannot use. The rules may include any or all of a set of mappings (e.g., used according to a set of lookup tables and / or databases), decision trees, equations, and / or any other components. Additionally or alternatively, the inputs may be processed using a set of models (e.g., deep learning models, machine learning models, pre-trained models, etc.), algorithms, and / or any other tools.

[0077]

[0083] In a preferred set of variants, processing the inputs and / or the set of features or parameters with the set of rules comprises comparing the location of the own agent with a set of one or more maps, which define which policies are available to the agent depending on its location. These policies available to the agent may be further refined (e.g. narrowed, expanded, etc.) and / or further curated in other ways based on other inputs or information described below (e.g. teleoperator's permission, time of day and / or situational information such as the agent's surroundings and / or weather, etc.).

[0078]

[0084] Additionally or alternatively, any other input (eg, as described above) may be compared to the map, such as directional information and / or any other information.

[0079]

[0085] The map is preferably a manually labeled map, and a set of policies is defined for each of a set of regions in the map (e.g., in a look-up table corresponding to the map). The set of policies defined for a region may optionally be determined based on any or all of the following: types of lanes in the region (e.g., inside lane, outside lane, leftmost lane of a four lane highway, rightmost lane of a four lane highway, left center lane of a four lane highway, right center lane of a four lane highway), lane markings associated with lanes in the region (e.g., solid lanes, dashed lanes, yellow lanes, white lanes, etc.), traffic signs associated with the region (e.g., traffic lights, traffic signs, stop signs, yield signs, etc.), traffic laws associated with the region (e.g., speed limits), road features associated with the region (e.g., road curvature, roundabouts, etc.), zones associated with the region (e.g., educational areas, hospital areas, residential areas, etc.), and / or any other features associated with the region.

[0080]

[0086] Additionally or alternatively, any or all of the policies associated with a map may be dynamically determined (e.g., using a set of pre-trained models).

[0081]

[0087] Additionally or alternatively, the set of policies may be determined based on the location and / or behavior of other vehicles in similar conditions (e.g., similar time of day, similar weather, etc.) In one variation, policies (e.g., actions, parameters associated with actions, etc.) may be selected based on, for example, how other vehicles are currently driving in the area and / or how other vehicles have driven in the area in the past.

[0082]

[0088] Determining the policies based on the map may optionally include implementing one or more geofencing processes to dynamically adjust the policies available to the agent in response to entering a region associated with a particular set of policies and based on the vehicle's location. Additionally or alternatively, any other process may be implemented to detect when the own agent is located within a particular region. Additionally or alternatively, a location may be determined for the own agent at each selection cycle, which location is then compared to the map and used to determine the set of available policies.

[0083]

[0089] The method may optionally include determining the set of maps and associated policy assignments based on historical information, based on a set of trained models, based on a labeling process, and / or based on any other process.

[0084]

[0090] S220 may optionally include processing any or all of the set of inputs to determine a situation awareness (also referred to herein equivalently as situation and / or environment awareness) associated with the own agent, which may be used to improve (e.g., add more available policies, remove policies, switch policies, etc.) and / or otherwise determine the set of policies available to the own agent and / or other agents or objects (e.g., as described below). The situation awareness is preferably associated with (e.g., determined based on) sensor information, more preferably at least sensor information from sensors onboard the own agent, but may additionally or alternatively be determined based on sensor information from infrastructure devices, sensor information collected at other agents, and / or any other sensor information. Additionally or alternatively, the agent's situation awareness may be determined based on sensor health information, environmental information (e.g., weather information, time of day, etc.), and / or any other information.

[0085]

[0091] In one variation, for example, the agent's situational awareness may indicate the agent's field of view, such as from any or all of: a set of cameras mounted on the own agent, a set of radar sensors mounted on the own agent, a set of lidar sensors mounted on the own agent, and / or any other sensors. Additionally or alternatively, the own agent's field of view may be determined based on (e.g., supplemented with) information collected from sensors (e.g., cameras) mounted on a set of infrastructure devices and / or based on any other information. In a first set of illustrative examples, for example, the agent's situational awareness may indicate that the agent's field of view is obstructed (e.g., by another agent or object, by glare from the sun, by heavy rain, by lack of light, etc.), which may then lead to improved policies available to the agent (e.g., as described below).

[0086]

[0092] In a second set of variations, in addition to or as an alternative to the first variation, the agent's situational awareness may be indicated based on sensor health information (e.g., sensor health state, sensor calibration state, sensor age, sensor operational state such as on / off state, etc.), where the sensor health information is associated with the agent's own sensors where the data collected from one or more sensors cannot be trusted (e.g., the sensor is malfunctioning, the sensor is broken, the sensor is not calibrated, etc.), which may subsequently result in improvements to the policies available to the agent (e.g., removing policies that require the sensor to run, removing policies that are dangerous to run without the sensor, etc.).

[0087]

[0093] In a third set of variations, additionally or alternatively to those described above, situational awareness may be determined based on environmental information associated with the own agent's environment, which may include, for example, any or all of the following: the time of day the own agent is traveling; features (e.g., based on historical information, based on a weather database, etc.) associated with the time of day (e.g., amount of light, ambient brightness, traffic volume, etc.); weather conditions (e.g., rain, snow, sleet, fog, overcast, sunny, etc.); traffic conditions (e.g., heavy traffic, light traffic, etc.); and / or any other information. Any or all of the environmental information may be dynamically determined (e.g., based on dynamically updated lookup tables, based on dynamically updated databases, based on third party websites and / or sensors and / or databases, based on sensor information, etc.), pre-determined (e.g., based on pre-defined lookup tables, data sets and / or sets of rules, etc.), and / or any combination.

[0088]

[0094] Additionally or alternatively, the own agent's situational awareness may be determined in other manners based on any suitable information.

[0089]

[0095] S220 may optionally additionally or alternatively further include processing the set of inputs to determine a scenario associated with the own agent, which may be used to refine and / or otherwise determine the set of policies available to the own agent and / or any other agents or objects (e.g., as described below). A scenario (also equivalently referred to herein as a scene and / or context) herein refers to a particular assessment (e.g., classification, categorization, assignment, etc.) of the agent's environment, such as an identification of the agent's particular environment. A scenario may indicate, for example, that the agent is within and / or proximate to a scene that may be difficult, non-standard, and / or somewhat dangerous for the own agent to navigate (e.g., approaching within a predefined distance threshold, within a predefined time, along a fixed path, etc.). Scenarios are preferably identified based on sensor information (e.g., from the own agent, from other agents, from infrastructure devices, etc.), such as any or all of the following: an uncertainty metric associated with the sensor information exceeding a predefined threshold (e.g., in a recognition module of the computing subsystem, in a prediction module of the computing subsystem, etc.); identification of one or more specific objects (e.g., unknown and / or rare objects such as construction equipment, a large number of pedestrians, a parade float, a large animal on the roadway, a traffic barricade, a smoke bomb, etc.) based on the sensor information; scenario classification of the sensor information using a classifier; human input (e.g., identification by a teleoperator, identification by an on-board operator, etc.); and / or any other information. Example scenarios include, but are not limited to, any or all of the following: construction zones (e.g., construction vehicles backing up), large numbers of pedestrians (e.g., gathering in the road for a protest, crossing the road in a large group, etc.), pedestrians on the roadway outside of a crosswalk, unexpected lane closures, broken / unpowered traffic lights, knocked over traffic signs, large animals in the roadway, objects on the road (e.g., shopping carts, parade floats, etc.), and / or any other scenarios.

[0090]

[0096] In a preferred set of variants, the scenarios associated with the own agent are determined using a classification module of the computing subsystem of the own agent, preferably comprising one or more classifiers (e.g. a perceptron classifier, a naive Bayes classifier, a decision tree classifier, a logistic regression classifier, a k-nearest neighbor classifier, a neural network classifier, an artificial neural network classifier, a deep learning classifier, a support vector machine classifier, etc.), more preferably one or more trained classifiers (e.g. a machine learning classifier, a deep learning classifier, etc.). Additionally or alternatively, the scenarios may be determined using one or more untrained classifiers, models (e.g. a machine learning model, a deep learning model, etc.), algorithms, decision trees and / or any other tool.

[0091]

[0097] For example, in a first set of examples, one or more streams of sensor information collected from at least one set of cameras mounted on the my agent are processed (e.g., at each selection cycle, in response to a trigger, etc.) using a classification module (e.g., a classifier, a set of classifiers, etc.) of the computing subsystem that detects whether a scenario exists. In additional or alternative examples, information from one or more infrastructure devices can be used to classify a scenario (e.g., using a classification module). In specific examples, this can function to detect that a scenario is ahead of the my agent (e.g., outside the current field of view of the my agent, along a fixed path and / or a currently planned path and / or a potential path for the my agent, etc.), which can function, for example, to reroute the my agent and / or otherwise improve the set of policies available to the my agent and / or otherwise facilitate other processes (e.g., requesting input from a teleoperator, selecting a particular policy associated with one or more objects in the my agent's environment, etc.).

[0092]

[0098] Additionally or alternatively, the scenario may be determined based on situational information, any other inputs (eg, as described in S210), and / or any other information.

[0093]

[0099] Additionally or alternatively, S220 may include any other suitable process.

[0094] 4.3 Method: Step S230 of determining a set of available policies based on a step of processing a set of inputs

[0100] Method 200 includes a step S230 of determining a set of available policies based on processing a set of inputs, which serves to determine and / or refine (e.g., iteratively refine) a subset of all policies that should be further considered by the autonomous agent in selecting a policy (e.g., a multi-policy decision-making process). S230 may additionally or alternatively serve to minimize and / or reduce the number of policies to be considered by the vehicle, thereby minimizing computing resources and / or latency associated with computing; maximize the number of simulations that can be run for each policy in the MPDM module (e.g., thereby ensuring that the least risky and / or otherwise optimal policy is ultimately selected); and / or provide any other benefit. Additionally or alternatively, S230 may perform any other suitable function.

[0095]

[0101] S230 is preferably performed in response to and based on S220, and optionally multiple times during operation of the own agent. Additionally or alternatively, S230 may be performed in response to another process of method 200, prior to S240 and / or any other process of method 200, in parallel with and / or as part of any other process of method 200, and / or at any other time.

[0096]

[0102] S230 preferably includes a step of referencing a set of one or more maps based on the agent location (e.g., dynamically determined using the sensor system described above) to determine at least a portion of the available policies. The maps preferably include a set of policy assignments corresponding to each of a set of regions defined in the set of maps. The regions are preferably defined by a set of boundaries (e.g., corresponding to geofenced boundaries, no geofenced boundaries, etc.), and the boundaries may optionally be defined based on one or more potential routes (e.g., predefined routes, dynamically determined routes) driven by the agent (e.g., for a self-agent operating as a shuttle, for a self-agent traveling a fixed delivery route, etc.). The regions may be any or all of independent (e.g., an agent is only in one region at a time), overlapping (e.g., an agent can be in multiple regions at a time, and the regions are determined per policy and overlap when multiple policies are available), and / or any combination. Additionally or alternatively, the regions may be defined without the self-agent's routes (e.g., for ride-sharing applications) and / or defined in other ways.

[0097]

[0103] In a preferred set of variations, one or more regions are determined based on, for example, the lane in which the ego-agent is traveling, and optionally, the surrounding lanes. In one variation, each region includes a single lane along a road segment (e.g., including a portion of a path segment, including the entire path, etc.), and the available policies for that region take into account the particular lane and the surrounding lanes. In a specific example, for example, if the region includes a right lane, right turns and left lane changes are included in the available policies, and in a region including a left lane, left turns and right lane changes are included in the available policies. Additionally, the region may include any number of road landmarks, such as traffic lights, traffic signs (e.g., stop signs, yield signs, one-way signs, speed limit signs, etc.), road features (e.g., speed bumps, potholes, highway shoulders, one-way roads, two-way roads, roundabouts, etc.), and the available policies are further determined (e.g., assigned) based on these road landmarks. Additionally or alternatively, the region may take into account zoning information (eg, educational districts, hospital districts, residential districts, business districts, etc.) and / or any other information.

[0098]

[0104] In one variation, a separate region is defined for each lane of a road segment (e.g., the entire fixed path, a portion of a fixed path, a modular component of a dynamically determined path, etc.), and the policy assigned to each region is determined, at least in part, based on the characteristics and surroundings of that lane. In specific examples, policies for turning into adjacent lanes are available in regions including any or all of one-way roads, inside lanes of multi-lane highways or two-way roads, left lanes in right lane regions, right lanes in left lane regions, and / or any other suitable regions.

[0099]

[0105] In variations involving infrastructure devices, the map assignment may optionally additionally or alternatively be determined based on proximity to and / or information collected by the infrastructure devices, the state / health of the infrastructure devices, and / or any other suitable information. In one example, for example, the set of policies available to the own agent at a location that is proximate to and / or viewable by a set of one or more infrastructure devices may potentially be larger than a similar location that is distant from and / or not viewable by the infrastructure devices.

[0100]

[0106] Additionally or alternatively, any or all of the boundaries may be determined based on pre-defined boundaries (eg, county boundaries, city-defined boundaries, etc.).

[0101]

[0107] The boundaries within the map are preferably statically determined (eg, during route planning), but may additionally or alternatively be dynamically determined.

[0102]

[0108] The policy assignment may optionally take into account a set of one or more fixed routes planned for the agent (e.g., in an implementation in which the autonomous agent operates as an autonomous shuttle traveling a fixed route). Additionally or alternatively, the autonomous agent may travel a non-fixed route (e.g., in an autonomous ride-sharing embodiment, an autonomous taxi embodiment, etc.).

[0103]

[0109] Additionally or alternatively, policy assignment may be determined based on aggregate and / or historical data, such as data regarding which policies have been executed by agents in the past and which policies were successful, failed, required on-board operator takeover, required teleoperator takeover, and / or any other suitable information.

[0104]

[0110] Additionally or alternatively, policy assignments may be determined (e.g., selected, improved, etc.) based on the behavior of other agents at the location and / or similar locations (e.g., same lane type, same neighborhood type, same speed limit, etc.) such that policies available to the own agent are similar to (e.g., mimic) the driving of other agents on the road. These other agents may refer to agents on the road with the own agent (e.g., surrounding the own agent), agents on the road before the own agent (e.g., based on historical information aggregated from a set of agents, based on historical information collected by the own agent, based on data collected at the same time and / or in the same situation, etc.), agents on a road similar to the own agent, and / or any other agents. In some examples, policies are selected and / or designed based on the behavior of other drivers at the location or similar locations (e.g., with a particular set of parameters such as speed limit and / or braking distance and / or acceleration).

[0105]

[0111] Each region of the map may optionally include any or all policy assignments that are the same in number as other regions, are below a predefined threshold (e.g., based on computing resources, based on latency requirements, etc.), are above a predefined threshold (e.g., include at least one policy, include at least two policies, etc.), have any suitable number of policies, and / or are otherwise determined.

[0106]

[0112] S230 may optionally further consider one or more situational features (e.g., as described above) when determining the policies available to the agent. The situational features are preferably determined based on the agent's situational awareness, which may be enabled by any or all of the agent's sensor systems, infrastructure device sensor systems, teleoperator input, on-board operator input, and / or any other suitable information. This may include, for example, awareness of any or all of: what other vehicle agents are surrounding (e.g., whether they are blocking the agent's view from one or more sensors, how the surrounding vehicles are driving, etc.), the environmental conditions surrounding or expected to surround the agent (e.g., lighting conditions, weather conditions, traffic conditions, time of day, etc.), and / or any other suitable environmental features (e.g., as described above).

[0107]

[0113] In one variation, for example, a first set of available policies is determined based on the map and the vehicle's position, and the available policies are further refined and / or expanded based on the agent's situational awareness. In a specific example, a subset of policies is excluded for consideration (e.g., in this selection cycle) based on adverse environmental conditions (e.g., bad weather, lack of sufficient light, etc.) and / or a set of policies is added to the set of available policies based on teleoperator input (e.g., selection of a high-risk policy for consideration by the agent's multi-policy decision-making module) and / or infrastructure devices (e.g., the agent has an obstructed view to ensure that the road ahead is clear). Additionally or alternatively, policies may be added and / or removed based on any other information.

[0108]

[0114] The policies associated with the situation are preferably determined based on a lookup table (e.g., a database), but may additionally or alternatively be determined based on a model, an algorithm, a set of rules, a decision tree, and / or any other tool. In a variant in which the first set of available policies is determined based on a lookup table associated with the map, the lookup table associated with the situation information may be any or all of: a part of the map lookup table, separate and distinct from the map lookup (e.g., and obtained based on the map lookup table), and / or otherwise related or independent of the map lookup table. In a set of illustrative examples, for example, the first set of policies for the own agent is selected based on the own agent's position in the map, and the first set of policies is reduced to a subset of the first set of policies (e.g., in the same lookup table, in a different lookup table, etc.) based on situation information that reduces the reliability of the own agent's perception of its environment. Additionally or alternatively, any or all of the policies may be determined using tools other than lookup tables (e.g., models, equations, decision trees, etc.).

[0109]

[0115] S230 may optionally additionally or alternatively include determining and / or refining a set of policies available to the own agent based on a scenario associated with the own agent. A particular scenario identified in the own agent's environment may be used, for example, to add policies to the set (e.g., add policies that help the own agent navigate the scenario, add policies that the own agent implements only in that particular scenario, add policies that violate traffic rules and / or conventions outside of that scenario, etc.), remove policies from the set, determine an entirely new set of policies, and / or otherwise determine policies available to the agent.

[0110]

[0116] S230 can optionally additionally or alternatively receive any type of user input that can be used to curate the policies available to the agent. In one variation, for example, the agent can receive input from a teleoperator located remotely from the agent to determine any or all of the available policies. Additionally or alternatively, non-human input can be received, such as from the output of a deep learning model stored on a remote computing system.

[0111]

[0117] In a preferred variant, the user input is used by the own agent to approve policies for consideration by the own agent (e.g., high-risk policies for which the system asks a human for permission), or in other words, to add policies to the set of available policies. Additionally or alternatively, the user input may be used to select a policy for the vehicle to select, to override a selected policy, to exclude a policy from consideration by the own agent (e.g., based on abnormal conditions), and / or in any other suitable manner.

[0112]

[0118] In one variation, for example, a request for user input from the teleoperator is triggered in response to detecting a scenario (e.g., any scenario, a specific scenario, etc.) associated with the own agent. In a specific example, the request for user input is in the form of a proposed policy that can only be considered by the own agent with approval from the teleoperator. Additionally or alternatively, the request for user input may include sensor data (e.g., a camera stream displayed by the teleoperator), proposed policies that the teleoperator excludes from consideration (e.g., if there are too many policies for the own agent to consider), and / or any other request.

[0113]

[0119] The available policies are preferably determined at S230 at least in part based on a predefined set of policies that are assigned (e.g., coded) to one or more maps (e.g., that take into account sensor values ​​to account for situational awareness) and / or lookup tables referenced by the maps. Additionally or alternatively, determining the policies may include any or all of: evaluating a set of rules, evaluating a decision tree, weighting any or all of a set of inputs (e.g., based on situational awareness, based on predefined weights, based on dynamically determined weights, using an algorithm, etc.), implementing one or more algorithms or models, implementing one or more probabilistic and / or statistical methods (e.g., Bayesian methods), and / or the available policies may be determined using any other tools.

[0114]

[0120] S230 may optionally include a step of determining (e.g., referencing from a look-up table, calculating, etc.) a risk associated with one or more policies, which may serve to reduce and / or prioritize the available policies. In one variation, for example, a risk is calculated and / or referenced for the policies, and policies with a risk above a predefined threshold are removed from the set of available policies. Additionally or alternatively, if the risk exceeds a predefined threshold, the teleoperator may be prompted to receive input to enable the policies whose risk exceeds the threshold.

[0115]

[0121] S230 may optionally include determining that there are too many policies available to the agent and / or that the risk of one or more policies is too high (e.g., based on computing resources, based on predefined thresholds, etc.), and one or more policies are then removed from consideration. Policies may be removed based on any or all of the calculated risk values, user input, historical data, a ranked list of policies, a random removal process, computing requirements associated with each policy, and / or based on any other characteristic.

[0116]

[0122] Additionally or alternatively, any or all of the policy assignments may be determined based on one or more machine learning models, deep learning models, predictive algorithms, and / or any other suitable tools.

[0117]

[0123] S230 may optionally include determining (e.g., predicting) a set of policies associated with any or all of a set of monitored objects in the own agent's environment. These policies may then be used, for example, in running a simulation associated with each of the sets of policies available according to the MPDM module (e.g., as described below). In a specific example, for example, assigning policies to each of a set of monitored objects in the own agent's environment (e.g., all objects perceivable by the own agent, all objects within a predefined distance threshold, etc.) may be part of an intent inference process used in the simulation of the MPDM module.

[0118]

[0124] Additionally or alternatively, intent inference may be performed without an assigned policy (e.g., using only position and / or velocity information associated with the monitored object), the MPDM module may be implemented in other suitable manners, the method may be performed without an MPDM module, and / or the method may be performed in other suitable manners.

[0119]

[0125] Additionally or alternatively, S230 may include any other suitable process.

[0120]

[0126] In a first variant, S230 includes the steps of: referencing a set of maps based on a dynamically determined location of the agent, the location specifying an area of ​​the map in which the agent is located; determining a set of policies available to the agent based on the area, the available policies may be any or all of: determined based on a rule-based approach, referenced from a look-up table, and / or determined in other manners; further curating the available policies based on the agent's situational awareness (e.g., based on the agent's sensor system, based on a sensor system of an infrastructure device, based on input of a teleoperator, etc.); further curating the available policies based on detection of specific scenarios in the agent's environment; optionally, reducing the number of available policies if the number of policies is too large; and optionally receiving input from the teleoperator regarding one or more available policies and / or high-risk policies currently unavailable to the agent, wherein the set of available policies may be further improved based on the teleoperator input.

[0121]

[0127] In a specific example (e.g., as shown in FIGS. 5A and 5B), the map may include any suitable region or combination of regions, such as any or all of regions A-G. The available policies for each region may be determined based on any or all of the following: referencing a lookup table (statistically assigning a set of policies based on vehicles located in the region; dynamically assigning a set of policies based on a particular location of the vehicles in the region and a set of rules; assigning a set of rules and / or algorithms, etc.); evaluating a set of rules and / or algorithms (e.g., based on a set of inputs including the location of the vehicles; hard-coding policies for each region of the map, etc.); evaluating a set of models (e.g., deep learning models); determining and / or adjusting the set of available policies based on the agent's environmental and / or situational awareness; facilitating, determining and / or updating the policies based on teleoperator input; and / or determining the available policies in any other suitable manner.

[0122]

[0128] In additional or alternative embodiments, S230 may include assigning policies to any or all of a set of monitored objects in the own agent's environment, the policies assigned to these monitored objects being preferably determined based on location, velocity and / or semantic (e.g., object classification) information associated with the monitored objects, and based on sensor information collected from the own agent and / or infrastructure devices.

[0123]

[0129] In a second variant, additionally or alternatively to the first variant, S230 includes refining a set of policies associated with the own agent's location (e.g., assigned to the map) based on situational awareness associated with the own agent, which is preferably associated with the own agent's sensors and / or other sensors associated with the system (e.g., sensors of infrastructure devices), but may additionally or alternatively be associated with one or more databases (e.g., third-party databases) and / or any other information source. In a specific example, a field of view associated with a camera and / or radar and / or lidar sensor mounted on the own agent is used to determine whether there is an obstruction in the own agent's field of view (e.g., another object blocking the field of view, weather conditions that reduce visibility, lighting conditions that reduce visibility, etc.), and if so, refine the policies based on the obstruction (e.g., remove policies that require this field of view). In additional or alternative examples, the time of day and / or other information (e.g., from a traffic database, a weather database, etc.) can be used to refine the policies (e.g., removing policies that are unsafe to execute at night, removing policies associated with speeds above a predefined threshold during times associated with heavy traffic, adding policies associated with speeds below a predefined threshold during times associated with heavy traffic).

[0124]

[0130] Additionally or alternatively, situational awareness may be used to determine policies in the absence of an agent's location and / or map, situational awareness may be used to obtain a particular map associated with the situational awareness, situational awareness may be determined based on any other information, situational awareness may be used to determine a new set of policies (e.g., with respect to policies associated with the location and / or map), and / or policies may be determined in other manners.

[0125]

[0131] In a third variant, in addition or as an alternative to the first and second variants, S230 includes a step of improving the set of policies (e.g., using a set of classifiers) based on scenarios detected in the environment of the own agent. In some cases, for example, a first set of policies may be determined based on the location of the own agent in a map, which is then optionally improved based on situational awareness associated with the own agent, and then improved (e.g., again) based on the scenario. Additionally or alternatively, the policies may be improved in another order (e.g., based on scenarios and then situational awareness, based on situational awareness and then improved based on location and then scenarios, etc.), determined without any of this information (e.g., without situational awareness, without location, without scenario, etc.), determined separately for each type of information (e.g., location vs. situational awareness vs. scenario), and then aggregated to determine the set of available policies (e.g., to include only those policies in the entire set), and / or the policies may be determined in other ways.

[0126]

[0132] In a specific example of the third variant (e.g., as shown in Fig. 6A-6C), a construction zone scenario is detected using a set of classifiers implemented in the computing subsystem and a set of sensor streams collected from sensors on board the own agent and / or from a set of infrastructure devices. For example, the classifiers may detect irregular behavior of the construction machine as the monitored object, such as a vehicle reversing on the roadway, a vehicle driving in a direction not along the lane, a vehicle oriented not along the lane, human movement, the presence of traffic cones, and / or any other indication associated with the scenario. Upon detecting the scenario, an alert is sent to the teleoperator and a policy is presented for the own agent that the teleoperator can accept and / or reject, and if the teleoperator accepts the policy, the policy is added to the set of policies available for the own agent to consider. Additionally or alternatively, the teleoperator may detect and / or classify the scenario and / or otherwise provide feedback associated with one or more policies.

[0127]

[0133] In another example, S230 may include, for example, determining and / or adjusting one or more policies assigned to a monitored object associated with a scenario available in the MPDM module (e.g., as described below). For example, in a construction zone scenario, a construction vehicle may be assigned a policy that provides for any or all of: more space for its movement (e.g., larger turns), a longer distance associated with backing up, a lower speed for its movement, and / or any other policies.

[0128]

[0134] In another example, a scenario is detected (e.g., using a set of classifiers) that indicates that a large group of pedestrians is present at a wide crosswalk. As a result of this scenario, a policy may be added to the set of available policies that allows the own agent to move through smaller gaps in people than would normally be allowed (e.g., without the crosswalk being completely clear), which can act to prevent the own agent from being stuck at the crosswalk for a significant amount of time (e.g., based on teleoperator input, without teleoperator input, etc.).

[0129]

[0135] Additionally or alternatively, S230 may include any other process and / or may be performed in any other manner as appropriate.

[0130] 4.4 Method: Implementing a Policy Step S240

[0136] Method 200 may optionally include a step S240 of implementing a policy operable to provide for the autonomous agent to implement the selected policy. S240 is preferably performed in response to and based on S230, and optionally multiple times during operation of the own agent. Additionally or alternatively, S240 may be performed in response to, prior to, in parallel with and / or during any other process of the method, or at any other time during method 200. Still additionally or alternatively, method 200 may be performed without S240.

[0131]

[0137] S240 preferably includes selecting a policy from the set of available policies determined in S230 using a multi-policy decision-making (MPDM) module, which evaluates each of the set of available policies (e.g., through a set of simulations) to select a policy to be implemented by the own agent (e.g., in the own agent's control system). The MPDM module preferably considers each of the set of available policies by simulating a set of one or more scenarios associated with each policy. Based on the simulations for each policy, a score or other metric may be determined for that policy, and the final selected policy is selected based on that score / metric (e.g., maximum score for safety, minimum score for risk, etc.). Additionally or alternatively, the MPDM module may evaluate the set of available policies in other ways.

[0132]

[0138] Additionally or alternatively, the set of available policies may be evaluated using any or all of a set of models (e.g., machine learning models, deep learning models, neural networks, etc.), algorithms, rule-based processes, decision trees, priority lists, other scores, and / or any other tools.

[0133]

[0139] In variants involving simulation of the own agent and monitored objects in its environment, the simulation is preferably performed based on an intent estimation process associated with the set of monitored objects, the intent estimation process functioning to predict intents associated with the monitored objects (e.g., current policies, future policies, a set of parameters such as speed and / or location associated with the current and / or future objects, other hypotheses, etc.). In one variant, for example, the MPDM module includes inferring and / or identifying intents of each of the agents surrounding the autonomous agent and / or surrounding the infrastructure device to generate one or more potential intents for these agents. The intent of the own agent in each simulation preferably corresponds to an available policy in the set of available policies corresponding to that simulation, and S240 includes running a set of simulations for each policy in the available set of policies, although it may additionally or alternatively be determined in other ways.

[0134]

[0140] In a specific example, for example, the MPDM module includes a simulator or similar machine or system that functions to estimate the impact of future (i.e., forward in time) behavioral policies (movements or actions) for each of the identified agents in the operating environment of the own agent along with each of a set of available policies executable by the autonomous agent. The simulation may provide data regarding, for example, position, motion information, and / or potential interactions (e.g., relative positions, relative velocities, relative accelerations, potential collisions, etc.) between the own agent and the monitored object.

[0135]

[0141] The simulation may additionally or alternatively include perturbing any or all of the set of inputs, such as perturbing predictive information (e.g., a set of intent and / or location and / or movement and / or semantic classification information) associated with the set of monitored objects, which serves to take into account various ways in which the monitored objects may behave, thereby allowing a safe policy to be selected for the own agent in light of any of these circumstances. For example, a policy may be selected for the own agent based on a collective set of simulations with a minimum risk (e.g., the average risk is minimum compared to the simulations of other policies, the maximum risk simulation is less than the maximum risk of other policies, etc.). Additionally or alternatively, any other characteristic (e.g., safety, time to destination, efficiency, minimum disruption to passengers on board the own agent, minimum disruption to other vehicles on the road, etc.) may be optimized.

[0136]

[0142] Specific examples of simulations are described in U.S. Application No. 16 / 514,624, filed July 17, 2019, and U.S. Application No. 15 / 923,577, filed March 16, 2018, each of which is incorporated by reference in its entirety into this specification.

[0137]

[0143] A set of simulations is preferably run for each policy of the set of available policies. The set of simulations may optionally include multiple simulations, preferably differing in which policies and / or intents are estimated for the set of monitored objects, but may additionally or alternatively differ in predicted parameters (e.g., position, speed, etc.) associated with the set of monitored objects, and / or differ in any other manner. In an illustrative example, each available policy is evaluated through the same number of simulations, but the policies may additionally or alternatively differ in the number of simulations run for each, may have the same time to run the simulations, may have different times to run the simulations, and / or may be evaluated in other ways. Each available policy in the set is preferably evaluated before selecting a policy. Additionally or alternatively, the policies may be evaluated in a prioritized manner (e.g., when the own agent has run out of time to select a policy) and / or may be evaluated in other ways.

[0138]

[0144] The step of selecting one of the set of available policies preferably functions to select one of the potential policies based on one or more predefined or dynamic selection criteria. The selection criteria may be based on any suitable policy selection factors that may be delineated prior to operation of the own agent, or may be dynamically based on one or more characteristics related to the operating environment or mode of operation of the own agent. For example, the selection criteria may be predetermined and / or set such that the own agent functions to select a policy that is most likely to be safely executed. In another example, if the operating environment of the autonomous vehicle includes emergency situations, the selection criteria may be dynamic and set such that the autonomous agent functions to select a behavior policy from a tractable set of behavior policies that require a (weighted) balance between efficiency of operation, safety, etc. Additionally or alternatively, the policy may be suitably selected in other ways.

[0139]

[0145] Additionally or alternatively, S240 may include and / or interact with any or all of the methods, processes, embodiments and examples as described in any or all of U.S. Application No. 16 / 514,624, filed July 17, 2019, and U.S. Application No. 15 / 923,577, filed March 16, 2018, each of which is incorporated herein by reference in its entirety, and may function to select an optimal policy based on predictive simulations using each of the available policies.

[0140]

[0146] In a first variation, S240 includes implementing a multi-policy decision-making process for selecting and implementing a policy from a set of available policies, as described above.

[0141]

[0147] Additionally or alternatively, S240 may be performed in other ways.

[0142] 5. Variations

[0148] In a first variation of method 200 (e.g., as shown in FIG. 7 ), the method includes the steps of: collecting a set of inputs at a sensor system of the agent, and optionally collecting the set of inputs from any or all of a roadside unit or other infrastructure device, a user (e.g., a teleoperator), and / or from any suitable source; processing the set of inputs (e.g., to determine parameters associated with the inputs); referencing a set of maps based on a dynamically determined location of the agent, the location identifying an area of ​​the map in which the agent is located; determining a set of policies available to the agent based on the area, the available policies may be any or all of: determined based on a rule-based approach, referenced from a look-up table, etc.; optionally, based on the agent's situational awareness, further curating the available policies based on the detected scenario (e.g., based on a sensor system of the agent, based on a sensor system of the infrastructure device, based on teleoperator input, etc.) and / or based on the detected scenario; optionally receiving teleoperator input regarding one or more available policies and / or high-risk policies that are not currently available to the agent (e.g., based on the detected scenario); optionally reducing the number of available policies (e.g., if the number of policies is too large); optionally processing each of the available policies with a multi-policy decision-making module of the agent to select a policy from the set of available policies; and implementing the policy selected from the set of available policies (e.g., based on the multi-policy decision-making module).

[0143]

[0149] In a second variant of method 200, in addition to or as an alternative to the first variant, the method includes the steps of: collecting a set of inputs at a sensor system of the agent, and optionally collecting the set of inputs from any or all of a roadside unit or other infrastructure device, a user (e.g., a teleoperator), and / or from any suitable source; processing the set of inputs (e.g., to determine parameters associated with the inputs); referencing a set of maps based on a dynamically determined location of the agent, the location identifying an area of ​​the map in which the agent is located; determining a set of policies available to the agent based on the area, the available policies may be any or all of: determined based on a rule-based approach, referenced from a lookup table, determined using a model, and / or determined in any other manner; The method may include any or all of the following steps: determining a set of weights associated with the set of policies based on the recognition (e.g., based on a sensor system of the agent, based on a sensor system of the infrastructure device, based on teleoperator input, etc.) and / or based on the detected scenario and / or teleoperator input; refining the set of policies based on the set of weights (e.g., eliminating policies associated with weights below a predefined threshold, ordering the policies based on their weights, etc.); optionally reducing the number of available policies (e.g., if the number of policies is too large); optionally processing each of the available policies using a multi-policy decision-making module of the agent to select a policy from the set of available policies; and implementing the policy selected from the set of available policies (e.g., based on the multi-policy decision-making module).

[0144]

[0150] In a third variant of method 200, in addition to or as an alternative to the first and second variants described above, the method includes the steps of: collecting a set of inputs at a sensor system of an agent, and optionally collecting the set of inputs from any or all of a roadside unit or other infrastructure device, a user (e.g. a teleoperator), and / or from any suitable source; processing the set of inputs (e.g. to determine parameters associated with the inputs); referencing a set of maps based on a dynamically determined location of the agent, the location identifying an area of ​​the map in which the agent is located; determining a set of policies available to the agent based on the area, the available policies may be any or all of: determined based on a rule-based approach, referenced from a look-up table; and optionally based on situational awareness of the agent (e.g. based on the agent's sensor system, the infrastructure device's further curating the available policies based on a sensor system, based on teleoperator input, etc. and / or based on input of the detected scenario; optionally receiving teleoperator input regarding one or more available policies and / or high risk policies that are not currently available to the agent (e.g., based on the detected scenario); optionally reducing the number of available policies (e.g., if the number of policies is too large); determining a set of predicted policies associated with any or all of the monitored objects in the agent's environment; optionally processing each of the available policies along with the monitored object's predicted policies using a multi-policy decision-making module to select a policy from the set of available policies for the agent; and implementing the selected policy from the set of available policies (e.g., based on the multi-policy decision-making module).

[0145]

[0151] Additionally or alternatively, method 200 may include any other suitable processes performed in any suitable order.

[0146]

[0152] Although omitted for the sake of brevity, preferred embodiments include all combinations and permutations of the various system components and various method processes, which may be performed in any suitable order, sequentially or simultaneously.

[0147]

[0153] As those skilled in the art will recognize from the foregoing detailed description and from the drawings and claims, modifications and variations can be made to the preferred embodiments of the invention without departing from the scope of the invention, which is defined in the following claims.

Claims

1. 1. A method for dynamically curating a set of policies available to an agent, the method comprising, at each selection step of a selection cycle: receiving a set of sensor inputs; determining a location of the agent; determining a set of policies based on the location using a labeled map; (a) obtaining a database based on the labeled map and the set of sensor inputs, and refining the set of policies based on the database; and (b) using a classification module to characterize a scenario of the agent based on the set of sensor inputs and further refine the set of policies based on the scenario, wherein in response to characterizing the scenario: Obtaining scenario-based policies; transmitting said scenario-based policy to a teleoperator for validation; and refining a set of policies, comprising: in response to receiving validation of the scenario-based policy from the teleoperator, including the scenario-based policy in the refined set of policies; refining the set of policies based on selecting a policy from the refined set of policies using a multi-policy decision module; implementing the policy in a control system of the agent; In response to receiving validation of the scenario-based policy from the teleoperator in a first selection step of the selection cycle, the method comprises the step of including the scenario-based policy in the set of improved policies; In response to receiving a rejection of the scenario-based policy, the method includes excluding the scenario-based policy from the set of refined policies.

2. 2. The method of claim 1, wherein refining the set of policies based on the database comprises eliminating a subset of the set of policies based on a field of view associated with at least one of the agent's set of cameras and a set of lidars.

3. The method of claim 1 , wherein the step of selecting a policy comprises the step of: running a set of multiple simulations for each policy of the set of policies with the multi-policy decision-making module.

4. 4. The method of claim 3, further comprising predicting a second set of policies associated with a set of monitored objects within the agent's environment, the set of multiple simulations being further performed based on the second set of policies.

5. 5. The method of claim 4, further comprising perturbing at least one of the second set of policies and a set of information associated with the set of surveillance objects, the set of information associated with the set of surveillance objects comprising at least one of a set of positions, a set of velocities, and a set of semantic data, and the set of simulations is further performed based on the perturbed second set of policies and at least one of the perturbed sets of information associated with the set of surveillance objects.

6. The step of refining the set of policies based on the scenario includes, in response to characterizing the scenario, comprising: obtaining a scenario-based policy, the scenario-based policy being separate and distinct from the set of policies; and transmitting the scenario-based policy to a teleoperator for validation.

7. 7. The method of claim 6, in response to receiving validation for the scenario-based policy from the teleoperator, the method further comprising the step of including the scenario-based policy in the set of improved policies.

8. 7. The method of claim 6, in response to receiving a rejection of the scenario-based policy, the method further comprising refraining from adding the scenario-based policy to the set of refined policies.

9. 2. The method of claim 1, further comprising: determining, for each set of monitored objects in the agent's environment, a second set of policies associated with the monitored agent, the multi-policy decision module being implemented based on the second set of policies.

10. The method of claim 9 , further comprising: performing an intent inference process on the supervised set of objects based on the second set of policies.

11. The method of claim 10 , further comprising: running a set of simulations based on a set of intents determined based on the intent estimation process.

12. The method of claim 1 , wherein the set of sensor inputs is further received from a set of infrastructure devices arranged in the agent's environment.

13. 1. A system for dynamically curating a set of policies available to an agent, the system comprising: a control subsystem of the agent; A sensor subsystem mounted on the agent; A labeled map and a computing subsystem in communication with the control subsystem, the teleoperator platform, the sensor subsystem, and the labeled map, the computing subsystem comprising: receiving a set of sensor inputs from the sensor subsystem; determining a location of the agent; determining a set of policies associated with the agent based on the location and the labeled map; (a) obtaining a database based on the labeled map and the set of sensor inputs, and refining the set of policies based on the database; and (b) using a classification module to characterize a scenario of the agent based on the set of sensor inputs and to refine the set of policies based on the scenario, wherein in response to characterizing the scenario: sending the scenario-based policy to a teleoperator of said teleoperator platform for validation; and refining the set of policies including including or excluding the scenario-based policy from the refined set of policies in response to the teleoperator validation received from the teleoperator platform; and refining the set of policies based on selecting a policy from the improved set of policies using a multi-policy decision module; a computing subsystem that implements the policy in a control system of the agent.

14. 14. The system of claim 13, further comprising a set of infrastructure devices arranged in the user's environment, the set of sensor inputs being further received from a second sensor subsystem disposed on board the set of infrastructure devices.

15. Refining the set of policies based on the database includes: eliminating a subset of the set of policies based on a field of view associated with at least one of a set of cameras located on the agent, a set of lidars located on the agent, and sensor data from a set of sensors located on infrastructure devices of the set of infrastructure devices; and refining a subset of the set of policies based on the sensor information from the set of sensors deployed on the infrastructure devices.

16. The system of claim 13 , wherein selecting the policy comprises running a set of multiple simulations for each policy of the set of policies with the multi-policy decision-making module.

17. 20. The system of claim 16, wherein the multi-policy decision-making module further determines a set of intents associated with a set of monitored objects in the agent's environment, and the set of the plurality of simulations is executed based on the set of intents.

18. 20. The system of claim 17, wherein the multi-policy decision-making module further perturbs at least one of the set of intent and a set of information associated with the set of surveillance objects, the set of information associated with the set of surveillance objects including at least one of a set of positions, a set of velocities, and a set of semantic data, and the set of simulations is further performed based on the perturbed set of intent and at least one of the perturbed set of information associated with the set of surveillance objects.

19. The computing subsystem further communicates with a teleoperator platform, and refining the set of policies based on the scenario comprises, in response to characterizing the scenario, obtaining a scenario-based policy, the scenario-based policy being separate and distinct from the set of policies; and transmitting the scenario-based policy to a teleoperator of the teleoperator platform for validation.

20. In response to receiving validation for the scenario-based policy from the teleoperator, the computing subsystem includes the scenario-based policy in the refined set of policies; and 20. The system of claim 19, wherein in response to receiving a rejection of the scenario-based policy, the computing subsystem refrains from adding the scenario-based policy to the improved set of policies.

Citation Information

Patent Citations

  • Multi-perspective system and method for behavioral policy selection by an autonomous agent

    US20200026286A1