Abstract: A method includes obtaining, by a first agent engine that generates actions for a first agent, a first objective of the first agent. In some implementations, the method includes generating, by the first agent engine, a first influence for a second agent engine that generates actions for a computer-generated reality (CGR) representation of a second agent. In some implementations, the first influence is based on the first objective of the first agent. In some implementations, the method includes triggering the CGR representation of the second agent to perform a set of one or more actions that advances the first objective of the first agent. In some implementations, the second agent engine generates the set of one or more actions based on the first influence generated by the first agent engine.

Abstract: In some implementations, a method of navigation mesh exploration is performed at a virtual agent operating system. The method includes: determining one or more first sensory perception regions for one or more senses of a virtual agent based on a first perceptual vector associated with the virtual agent; generating a first portion of a navigation mesh for the XR environment based on the one or more first sensory perception regions, wherein the first portion of the navigation mesh includes candidate subsequent locations different from the first location; and in response to detecting movement of the virtual agent to a respective candidate subsequent location among candidate subsequent locations, generating a second portion of the navigation mesh for the XR environment based on one or more second sensory perception regions for the one or more senses of the virtual agent relative to the respective candidate subsequent location.

Abstract: An encoded artificial intelligence (AI) behavior specification is received. A data generation configuration specification is received. And a deep neural network configuration specification is received. A training data set based on the data generation configuration specification is generated. An AI behavior deep neural network that conforms to the deep neural network configuration specification is trained using at least a subset of the generated training data. The trained AI behavior deep neural network is provided from a remote AI add-in service to a development environment.

Abstract: In one implementation, a method for improved motion planning. The method includes: obtaining a macro task for a virtual agent within a virtual environment; generating a search-tree based on at least one of the macro task, a state of the virtual environment, and a state of the virtual agent, wherein the search-tree includes a plurality of task nodes corresponding to potential tasks for performance by the virtual agent in furtherance of the macro task; and determining physical motion plans (PMPs) for at least some of the plurality of task nodes within the search-tree in order to generate a lookahead planning gradient for the first time, wherein a granularity of a PMP for a respective task node in the first search-tree is a function of the temporal distance of the respective task node from the first time.

Abstract: A method includes obtaining, by a first agent engine that generates actions for a first agent, a first objective of the first agent. In some implementations, the method includes generating, by the first agent engine, a first influence for a second agent engine that generates actions for a computer-generated reality (CGR) representation of a second agent. In some implementations, the first influence is based on the first objective of the first agent. In some implementations, the method includes triggering the CGR representation of the second agent to perform a set of one or more actions that advances the first objective of the first agent. In some implementations, the second agent engine generates the set of one or more actions based on the first influence generated by the first agent engine.

Abstract: Various implementations disclosed herein include devices, systems, and methods for controlling motion of CGR objects. In various implementations, a device includes a non-transitory memory and one or more processors coupled with the non-transitory memory. In some implementations, a method includes determining, by a first animation controller, values for a first set of animation parameters associated with a first animation for a computer-generated reality (CGR) object. In some implementations, the CGR object is associated with a plurality of joints. In some implementations, the method includes generating, by a motion controller, respective joint movements for the plurality of joints based on the values for the first set of animation parameters. In some implementations, the method includes manipulating the CGR object in accordance with the respective joint movements for the plurality of joints in order to provide an appearance that CGR object is moving within a degree of similarity to the first animation.

Abstract: In some implementations, a method of improved pathfinding is performed at a virtual agent operating system including non-transitory memory and one or more processors coupled with the non-transitory memory. The method includes: determining an initial path for a virtual agent to a target destination based at least in part on a navigation mesh of an XR environment; actuating locomotive elements of the virtual agent in order to move the virtual agent according to the initial path; while moving according to the initial path, detecting a node of a navigation graph; in response to detecting the node of the navigation graph: obtaining navigation information from the node of the navigation graph; and determining an updated path from the node to the target destination based at least in part on the navigation mesh and the navigation information; and actuating the locomotive elements of the virtual agent according to the updated path.

Abstract: A method includes maintaining a hierarchical datastore for an agent instantiated in a computer-generated reality (CGR) environment. The hierarchical datastore includes a first storage hierarchy associated with a first data type and a second storage hierarchy associated with a second data type. The method includes detecting a sensory input that includes sensory input data. The method includes determining a type of the sensory input data. The method includes, in response to the sensory input data being of the first data type, storing the sensory input data in the first storage hierarchy for a first amount of time associated with the first storage hierarchy. The method includes, in response to the sensory input data being of the second data type, storing the sensory input data in the second storage hierarchy for a second amount of time that is different from the first amount of time.

Abstract: An encoded artificial intelligence (AI) behavior specification is received. A data generation configuration specification is received. And a deep neural network configuration specification is received. A training data set based on the data generation configuration specification is generated. An AI behavior deep neural network that conforms to the deep neural network configuration specification is trained using at least a subset of the generated training data. The trained AI behavior deep neural network is provided from a remote AI add-in service to a development environment.

Abstract: A specification of a problem using an artificial intelligence planning language is received. Machine learning features are determined using a computer processor and the specification of the problem. Using a trained machine learning model that is trained to approximate an automated planner and the determined machine learning features, a machine learning model result is determined. An action to perform is determined based on the machine learning model result.

Abstract: One or more sensory inputs of an autonomous artificial intelligence (AI) computer character are received. One or more beliefs of the autonomous AI computer character are determined using a computer processor and one or more goals of the autonomous AI computer character are identified. A machine learning model is used to automatically determine an action of the autonomous AI computer character based at least in part on the one or more sensory inputs, the one or more beliefs, and the one or more goals of the autonomous AI computer character.