University of Massachusetts

Nissan North America, Inc.

Renault S.a.s.

United States of America As Represented by the Administrator of Nasa

Shlomo Zilberstein

Stefan Witwicki

Kyle Hollins Wray

Liam Pedersen

Najamuddin Mirza Baig

Omar Bentahar

Astha Vagadia

Arec Jamgochian

Connor Basich

Christopher Ostafew

Julius S Gyorfi

Melissa Cefkin

Viju James

Laura Cesafsky

Sparsh Sharma

Justin Svegliato

A first method includes detecting, based on sensor data, an environment state; selecting an action based on the environment state; determining an autonomy level associated with the environment state and the action; and performing the action according to the autonomy level. The autonomy level can be selected based at least on an autonomy model and a feedback model. A second method includes calculating, by solving an extended Stochastic Shortest Path (SSP) problem, a policy for solving a task. The policy can map environment states and autonomy levels to actions and autonomy levels. Calculating the policy can include generating plans that operate across multiple levels of autonomy.

Online state space refinement for competence-aware systems

A first distinct vehicle operational scenario is identified for an autonomous vehicle (AV). A first set of candidate vehicle control actions are received from a model that provides a first solution to the first distinct vehicle operational scenario. An action is selected from the first set of candidate vehicle control actions. The AV is controlled based on the action. The first solution is obtained offline in a first idealized situation that is decoupled from a current context of the AV.

Risk-aware executor with action set recommendations

Lane-level route planning includes obtaining lane-level information of a road, where the road includes a first lane and a second lane and the lane-level information includes first lane information related to the first lane and second lane information related to the second lane; converting the lane-level information to probabilities for a state transition function; receiving a destination; and obtaining a policy as a solution to a model that uses the state transition function.

Learning in lane-level route planner

Providing explanations in route planning includes determining a route based on at least two objectives received from a user, where a second objective of the at least two objectives is constrained to within a slack value of a first objective of the at least two objectives; receiving, from the user, a request for an explanation as to an action along the route; and providing the explanation to the user. The explanation describes an extent of violating the slack value.

Explainability and interface design for lane-level route planner

Real-time decision-making for a vehicle using belief state determination is described. Operational environment data is received while the vehicle is traversing a vehicle transportation network, where the data includes data associated with an external object. An operational environment monitor establishes an observation that relates the object to a distinct vehicle operation scenario. A belief state model of the monitor computes a belief state for the observation directly from the operational environment data. The monitor provides the computed belief state to a decision component implementing a policy that maps a respective belief state for the object within the distinct vehicle operation scenario to a respective candidate vehicle control action. A candidate vehicle control action is received from the policy of the decision component, and a vehicle control action is selected for traversing the vehicle transportation from any available candidate vehicle control actions.

Belief state determination for real-time decision-making

An autonomous vehicle traverses a vehicle transportation network using a multi-objective policy based on a model for specific scenarios. The multi-objective policy includes a topographical map that shows a relationship between at least two objectives. The autonomous vehicle receives a candidate vehicle control action associated with each of the at least two objectives. The autonomous vehicle selects a vehicle control action based on a buffer value that is associated with the at least two objectives. The autonomous vehicle traverses a portion of the vehicle transportation network in accordance with the selected vehicle control action.

Objective-based reasoning in autonomous vehicle decision-making

Traversing, by an autonomous vehicle, a vehicle transportation network, may include identifying a distinct vehicle operational scenario, wherein traversing the vehicle transportation network includes traversing a portion of the vehicle transportation network that includes the distinct vehicle operational scenario, communicating shared scenario-specific operational control management data associated with the distinct vehicle operational scenario with an external shared scenario-specific operational control management system, operating a scenario-specific operational control evaluation module instance including an instance of a scenario-specific operational control evaluation model of the distinct vehicle operational scenario, and wherein operating the scenario-specific operational control evaluation module instance includes identifying a policy for the scenario-specific operational control evaluation model, receiving a candidate vehicle control action from the policy for the scenario-specific operational control evaluation model, and traversing a portion of the vehicle transportation network based on the candidate vehicle control action.

Shared autonomous vehicle operational management

A processor is configured to execute instructions stored in a memory to identify distinct vehicle operational scenarios; instantiate decision components, where each of the decision components is an instance of a respective decision problem, and where the each of the decision components maintains a respective state describing the respective vehicle operational scenario; receive respective candidate vehicle control actions from the decision components; select an action from the respective candidate vehicle control actions, where the action is from a selected decision component of the decision components, and where the action is used to control the AV to traverse a portion of the vehicle transportation network; and generate an explanation as to why the action was selected, where the explanation includes respective descriptors of the action, the selected decision component, and a state factor of the respective state of the selected decision component.

Explainability of autonomous vehicle decision making

Autonomous vehicle operation with explicit occlusion reasoning may include traversing, by a vehicle, a vehicle transporation network. Traversing the vehicle transportation network can include receiving, from a sensor of the vehicle, sensor data for a portion of a vehicle operational environment, determining, using the sensor data, a visibility grid comprising coordinates forming an unobserved region within a defined distance from the vehicle, computing a probability of a presence of an external object within the unobserved region by comparing the visibility grid to a map (e.g., a high-definition map), and traversing a portion of the vehicle transportation network using the probability. An apparatus and a vehicle are also described.

Autonomous vehicle operation with explicit occlusion reasoning

A method for use in traversing a vehicle transportation network by an autonomous vehicle (AV) includes traversing, by the AV, the vehicle transportation network. Traversing the vehicle transportation network includes identifying a distinct vehicle operational scenario; instantiating a first decision component instance; receiving a first set of candidate vehicle control actions from the first decision component instance; selecting an action; and controlling the AV to traverse a portion of the vehicle transportation network based on the action. The first decision component instance is an instance of a first decision component modeling the distinct vehicle operational scenario.

Risk aware executor with action set recommendations

Traversing a vehicle transportation network includes operating a scenario-specific operational control evaluation module instance. The scenario-specific operational control evaluation module instance includes an instance of a scenario-specific operational control evaluation model of a distinct vehicle operational scenario. Operating the scenario-specific operational control evaluation module instance includes identifying a multi-objective policy for the scenario-specific operational control evaluation model. The multi-objective policy may include a relationship between at least two objectives. Traversing the vehicle transportation network includes receiving a candidate vehicle control action associated with each of the at least two objectives. Traversing the vehicle transportation network includes selecting a vehicle control action based on a buffer value. Traversing the vehicle transportation network includes performing the selected vehicle control action, determining a preference indicator for each objective, and updating the multi-objective policy.

Learning safety and human-centered constraints in autonomous vehicles

A processor is configured to execute instructions stored in a memory to determine, in response to identifying vehicle operational scenarios of a scene, an action for controlling the AV, where the action is from a selected decision component that determined the action based on level of certainty associated with a state factor; generate an explanation as to why the action was selected, such that the explanation includes respective descriptors of the action, the selected decision component, and the state factor; and display the explanation in a graphical view that includes a first graphical indicator of a world object of the selected decision component, a second graphical indicator describing the state factor, and a third graphical indicator describing the action.

Autonomous vehicle operational management may include traversing, by an autonomous vehicle, a vehicle transportation network. Traversing the vehicle transportation network may include operating a scenario-specific operational control evaluation module instance, wherein the scenario-specific operational control evaluation module instance is an instance of a scenario-specific operational control evaluation module, wherein the scenario-specific operational control evaluation module implements a partially observable Markov decision process. Traversing the vehicle transportation network may include receiving a candidate vehicle control action from the scenario-specific operational control evaluation module instance, and traversing a portion of the vehicle transportation network based on the candidate vehicle control action.

Autonomous vehicle operational management including operating a partially observable Markov decision process model instance

Introspective competence modeling for AV decision making

A vehicle traversing a vehicle transportation network may use a scenario-specific operational control evaluation model instance. A multi-objective policy for the model is received, wherein the policy includes at least a first objective, a second objective, and a priority of the first objective relative to the second objective. A representation of the policy (e.g., the first objective, the second objective, and the priority) is generated using a user interface. Based on feedback to the user interface, a change to the multi-objective policy for the scenario-specific operational control evaluation model is received. The change is to the first objective, the second objective, the priority, of some combination thereof. Then, for determining a vehicle control action for traversing the vehicle transportation network, an updated multi-objective policy for the scenario-specific operational control evaluation model is generated to include the change to the policy.

Growing community of inventors

Amherst, MA, United States of America

Shlomo Zilberstein