Granted Patent B2

US 12,505,346 · App. 18/292,165 · Granted Dec 23, 2025

Inventors: Scott Ellison Reed (Atlanta, GA); Konrad Zolna (London, GB); Emilio Parisotto (London, GB); Tom Erez (London, GB); Alexander Novikov (London, GB); Jack William Rae (London, GB); Misha Man Ray Denil (London, GB); Joao Ferdinando Gomes de Freitas (London, GB); Oriol Vinyals (London, GB); Sergio Gomez (London, GB); Ashley Deloris Edwards (London, GB); Jacob Bruce (London, GB); Gabriel Barth-Maron (London, GB)

Assignee: GDM Holding LLC

G06N3/08G06N3/04

Abstract

Claims (65)

1 . A method performed by one or more computers for selecting actions to be performed by an agent to interact with an environment using an action selection neural network, wherein the agent performs an action at each control step in a sequence of control steps, the method comprising, at each control step in the sequence of control steps:

generating a current representation of a state of a task being performed by the agent in the environment as of the current control step as a sequence of data elements;

autoregressively generating a sequence of data elements that collectively define a current action to be performed by the agent at the current control step, wherein the sequence of data elements defining the current action comprises a respective data element at each of a plurality of positions and each respective data element represents a respective control dimension of the current action to be performed at the control step, and wherein the autoregressive generation comprises, for each of one or more positions starting from a first position in the sequence of data elements defining the current action:

processing the current representation of the state of the task using the action selection neural network to generate a score distribution over a set of possible data elements;

selecting a data element for the position in the sequence of data elements representing the current action in accordance with the score distribution, the selecting comprising at least one of selecting a data element with a highest score under the score distribution, or sampling a data element in accordance with a probability distribution over the set of possible data elements that is defined by the score distribution over the set of possible data elements;

updating the current representation of the state of the task by concatenating the selected data element for the position to the current representation of the state of the task; and

providing the current representation of the state of the task for subsequent processing by the action selection neural network to generate a data element at a next position in the sequence of data elements defining the current action to be performed at the control step;

after autoregressively generating the sequence of data elements representing the current action, causing the agent to perform the current action at the current control step; and

proceeding to a next control step.

2 . The method of claim 1 , wherein for each control step in the sequence of control steps, generating the current representation of the state of the task as of the current control step comprises:

receiving a current observation characterizing a state of the environment at the current control step, wherein the current observation is defined by a collection of numerical values;

generating a representation of the current observation as a sequence of data elements, wherein generating the representation of the current observation as a sequence of data elements comprises concatenating each numerical value in the collection of numerical values defining the current observation into a sequence of numerical values in a predefined order, and optionally discretizing each numerical value in the collection of numerical values defining the current observation; and

including the representation of the current observation as a sequence of data elements in the current representation of the state of the task as of the current control step.

3 . The method of claim 2 , wherein the current observation characterizing the current state of the environment at the current control step comprises an image defined by an array of pixels.

4 . The method of claim 2 , wherein generating the representation of the current observation as a sequence of data elements comprises:

combining a target return to be achieved by interaction of the agent with the environment with the representation of the current observation as a sequence of data elements, wherein the target return defines a cumulative measure of rewards to be achieved as a result of the interaction of the agent with the environment.

5 . The method of claim 2 , wherein for each control step after a first control step in the sequence of control steps, including the representation of the current observation as a sequence of data elements in the current representation of the state of the task as of the current control step comprises:

receiving a representation of the state of the task as of a previous control step as a sequence of data elements; and

concatenating the representation of the current observation as a sequence of data elements to the representation of the state of the task as of the previous control step as a sequence of data elements to generate the current representation of the state of the task as of the current control step.

6 . The method of claim 5 , wherein the representation of the state of the task as of the previous control step represents, for each control step preceding the current control step: (i) a respective observation characterizing a state of the environment at the control step, and (ii) a respective action performed by the agent at the control step.

7 . The method of claim 2 , wherein at a first control step in the sequence of control steps, including the representation of the current observation as a sequence of data elements in the current representation of the state of the task as of the current control step comprises:

receiving a prompt that comprises data characterizing the task to be performed by the agent in the environment;

generating a representation of the prompt as a sequence of data elements; and

concatenating the representation of the current observation as a sequence of data elements to the representation of the prompt as a sequence of data elements to generate the current representation of the state of the task as of the current control step.

8 . The method of claim 7 , wherein the prompt comprises one or more of: a demonstration of the task, a goal observation characterizing a goal state of the environment, or a sequence of text in a natural language that provides instructions related to the task.

9 . The method of claim 2 , wherein the current observation comprises an image, and wherein generating the representation of the current observation as a sequence of data elements comprises:

generating a respective initial patch embedding corresponding to each of a plurality of patches in the image;

processing the initial patch embeddings using an encoder neural network to generate a respective final patch embedding for each of the plurality of patches in the image;

wherein each final patch embedding is included as a respective data element in the sequence of data elements representing the current observation.

10 . The method of claim 9 , wherein generating a respective initial patch embedding corresponding to a patch in the image comprises:

generating a pixel embedding representing pixels in the patch in the image;

generating a patch positional embedding representing a position of the patch in the image; and

generating the initial patch embedding for the patch by combining the pixel embedding and the patch positional embedding for the patch.

11 . The method of claim 1 , wherein the action selection neural network has been trained on a set of training examples, wherein for each training example:

the training example is represented as a sequence of data elements;

at least one of the data elements in the sequence of data elements representing the training example is designated as an action data element; and

training the action selection neural network on the training example comprises training the action selection neural network to generate the action data elements included in the training example.

12 . The method of claim 11 , wherein the set of training examples includes respective training examples from multiple different control domains, wherein each control domain is associated with: (i) a corresponding agent, (ii) a corresponding environment, and (iii) a corresponding task, wherein each training example from each control domain characterizes interaction of the corresponding agent with the corresponding environment by performing actions to accomplish the corresponding task.

13 . The method of claim 12 , wherein the multiple different control domains include a first control domain where observations of the corresponding environment have a first dimensionality, and a second control domain where observations of the corresponding environment have a second, different dimensionality.

14 . The method of claim 12 , wherein the multiple different control domains include a first control domain where actions performed by the corresponding agent have a first dimensionality, and a second control domain where actions performed by the corresponding agent have a second, different dimensionality.

15 . The method of claim 11 , wherein the set of training examples includes a plurality of language modeling training examples, wherein each language modeling training example represents a sequence of text in a natural language.

16 . The method of claim 1 , wherein the action selection neural network comprises a plurality of self-attention neural network layers.

17 . The method of claim 1 , wherein for each position starting from the first position in the sequence of data elements representing the current action, selecting the data element for the position comprises:

selecting a data element having a highest score under the score distribution.

18 . The method of claim 1 , wherein for each control step in the sequence of control steps, the sequence of data elements representing the state of the task as of the current control step comprises: a sequence of numerical values; a sequence of ordered collections of numerical values; or a sequence that includes numerical values at some positions and ordered collections of numerical values at other positions; wherein the ordered collections of numerical values comprise.

19 . A system comprising:

one or more computers; and

one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for selecting actions to be performed by an agent to interact with an environment using an action selection neural network, wherein the agent performs an action at each control step in a sequence of control steps, the operations comprising, at each control step in the sequence of control steps:

generating a current representation of a state of a task being performed by the agent in the environment as of the current control step as a sequence of data elements;

autoregressively generating a sequence of data elements that collectively define a current action to be performed by the agent at the current control step, wherein the sequence of data elements defining the current action comprises a respective data element at each of a plurality of positions and each respective data element represents a respective control dimension of the current action to be performed at the control step, and wherein the autoregressive generation comprises, for each of one or more positions starting from a first position in the sequence of data elements defining the current action:

processing the current representation of the state of the task using the action selection neural network to generate a score distribution over a set of possible data elements;

selecting a data element for the position in the sequence of data elements representing the current action in accordance with the score distribution, the selecting comprising at least one of selecting a data element with a highest score under the score distribution, or sampling a data element in accordance with a probability distribution over the set of possible data elements that is defined by the score distribution over the set of possible data elements;

updating the current representation of the state of the task by concatenating the selected data element for the position to the current representation of the state of the task; and

providing the current representation of the state of the task for subsequent processing by the action selection neural network to generate a data element at a next position in the sequence of data elements defining the current action to be performed at the control step;

after autoregressively generating the sequence of data elements representing the current action, causing the agent to perform the current action at the current control step; and

proceeding to a next control step.

20 . One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for selecting actions to be performed by an agent to interact with an environment using an action selection neural network, wherein the agent performs an action at each control step in a sequence of control steps, the operations comprising, at each control step in the sequence of control steps:

generating a current representation of a state of a task being performed by the agent in the environment as of the current control step as a sequence of data elements;

autoregressively generating a sequence of data elements that collectively define a current action to be performed by the agent at the current control step, wherein the sequence of data elements defining the current action comprises a respective data element at each of a plurality of positions and each respective data element represents a respective control dimension of the current action to be performed at the control step, and wherein the autoregressive generation comprises, for each of one or more positions starting from a first position in the sequence of data elements defining the current action:

processing the current representation of the state of the task using the action selection neural network to generate a score distribution over a set of possible data elements;

selecting a data element for the position in the sequence of data elements representing the current action in accordance with the score distribution, the selecting comprising at least one of selecting a data element with a highest score under the score distribution, or sampling a data element in accordance with a probability distribution over the set of possible data elements that is defined by the score distribution over the set of possible data elements;

updating the current representation of the state of the task by concatenating the selected data element for the position to the current representation of the state of the task; and

providing the current representation of the state of the task for subsequent processing by the action selection neural network to generate a data element at a next position in the sequence of data elements defining the current action to be performed at the control step;

after autoregressively generating the sequence of data elements representing the current action, causing the agent to perform the current action at the current control step; and

proceeding to a next control step.

Assignments (2)

Continuity (3)

References Cited (125)