DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Drawings
The drawings are objected to because: In FIG. 6, “Fully-Constrained Controller 406” should read “Fully-Constrained Controller 404”. Corrected drawing sheets in compliance with 37 CFR 1.121(d) are required in reply to the Office action to avoid abandonment of the application. Any amended replacement drawing sheet should include all of the figures appearing on the immediate prior version of the sheet, even if only one figure is being amended. The figure or figure number of an amended drawing should not be labeled as “amended.” If a drawing figure is to be canceled, the appropriate figure must be removed from the replacement sheet, and where necessary, the remaining figures must be renumbered and appropriate changes made to the brief description of the several views of the drawings for consistency. Additional replacement sheets may be necessary to show the renumbering of the remaining figures. Each drawing sheet submitted after the filing date of an application must be labeled in the top margin as either “Replacement Sheet” or “New Sheet” pursuant to 37 CFR 1.121(d). If the changes are not accepted by the examiner, the applicant will be notified and informed of any required corrective action in the next Office action. The objection to the drawings will not be held in abeyance.
Specification
The disclosure is objected to because of the following informalities:
In page 37, line 20, “the path 1112” should read “the path 1102”.
Appropriate correction is required.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 1, 6, 8, 11, 14, 18-20 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Bashkirov et al. (US 2022/0383576 A1, hereinafter “Bashkirov”).
Regarding claim 1, Bashkirov teaches A computer-implemented method for animating characters, the method comprising: (page 2, para. [0024], “a computer animation method may use target segmentation masks for multiple camera views of a character to resolve ambiguities in pose”).
receiving one or more goals specified in one or more modalities; (page 6, para. [0048], “The input may include any or all of the following features: goal orientations, joint sensor readings, action at previous time step, actuator inputs at previous time step, gravity vector in local reference frame, accelerometer readings, gyro readings, and foot pressure sensor readings.”)
generating, via a trained machine learning model and based on the one or more goals, a first action for a character to perform, wherein the trained machine learning model is trained to process inputs in multiple modalities; and (page 5, para. [0045], “From the chosen candidate poses a robot may be controlled using the Neural Networks 420. Outputs of the Neural Networks 420 include an action 425 and a value 427. The action 425 corresponds to the control inputs to the robot 415. The value 427 is an internal training algorithm quantity. It is needed only during training step and is used to estimate the effect of random attempts at improvement”; page 5, para. [0044], “The Neural Networks 420 may be trained to determine the next two poses from a current pose. The Neural Networks 420 training may include the use of a character model in a physics simulation. Motion capture or hand animated poses may be used as a target and the Neural Network 420 may trained to replicate the target poses within the constraints of the physics simulation using a machine learning algorithm”). Note that: (1) an action as a first action can be output or generated from the trained neural network model 420; and (2) the trained neural network model 420 are trained to replicate or generate target poses by feeding the corresponding goal inputs above as the inputs to the neural network model.
causing the character to perform the first action within a computer-based (page 1, para. [0004], “The computer uses the model to compute the exact position and orientation of that certain character, which is eventually rendered into an image. Thus by changing the animation variable values over time, the animator creates motion by making the character move from frame to frame”). Note that: the computer computes pose of the character to render an image in the computer-based environment. or physical environment. (page 5, para. [0041], “the animation program may use an output of the pose optimization process at 408 to generate robot control inputs 413, as indicated at 414. The animation program may supply the control inputs 413 to a robot controller 415, which converts the control inputs to control signals that are transmitted to an articulated robot 417”; FIG. 4A: “Robot Controller 415” generates the second action of an articulated robot “417”). Note that: the robot controller can control the robot in the physical world using the control inputs from the animation program to perform the action.
Regarding claim 6, Bashkirov teaches The computer-implemented method of claim 1, wherein the one or more goals include at least one of a set of constraints associated with a subset of joints belonging to the character for one or more frames, a textual description, or an object for the character to interact with. (page 4, para. [0039], “The opt1m1zation process 408 makes the segmentation mask for a frame of the sequence 401 match a corresponding projection of a candidate pose of the model 405 and at the same time makes sure that the movement of the animated character is physically consistent, e.g., doesn't cause the animation character to fall or violate joint constraints”). Note that: joint constraints can be regarded as a set of constraints associated with a subset of joints belonging to the character for one or more frames.
Regarding claim 8, Bashkirov teaches The computer-implemented method of claim 1, further comprising:
generating, via the trained machine learning model and based on the one or more goals, a second action for the character to perform subsequent to the first action; and (Bashkirov, FIG. 4A: a loop of “Next Frame 412” -> “401” -> “404” -> “405” –> “406” –> “407” -> “408” -> “Optimum pose 409” -> 410” -> “411” –> “Next Frame 412” is formed to generate a subsequent pose (action) or a second action of current pose (the first action)).
causing the character to perform the second action within the computer-based (page 1, para. [0004], “The computer uses the model to compute the exact position and orientation of that certain character, which is eventually rendered into an image. Thus by changing the animation variable values over time, the animator creates motion by making the character move from frame to frame”). Note that: the computer computes the second pose (action) of the character to render an image in the computer-based environment. or physical environment. (page 5, para. [0041], “the animation program may use an output of the pose optimization process at 408 to generate robot control inputs 413, as indicated at 414. The animation program may supply the control inputs 413 to a robot controller 415, which converts the control inputs to control signals that are transmitted to an articulated robot 417”; FIG. 4A: “Robot Controller 415” generates the second action of an articulated robot “417”). Note that: the robot controller can control the robot in the physical world / environment using the control inputs from the animation program to perform the action.
Claim 11 reciting “One or more non-transitory computer-readable media storing instructions that, when executed by at least one processor, cause the at least one processor to perform the steps of:” is corresponding to the method of claim 1. Therefore, claim 11 is rejected with the same prior art and citations for claim 1.
In addition, Bashkirov teaches One or more non-transitory computer-readable media storing instructions that, when executed by at least one processor, cause the at least one processor to perform the steps of: (page 8, claim 28, “A non-transitory computer readable medium having executable instructions embodied therein that, when executed by a computer cause the computer to implement a method for computer animation, the method comprising”; page 6, para. [0053], “The processor unit 503 may execute one or more programs 517, portions of which may be stored in the memory 504 and the processor 503 may be operatively coupled to the memory, e.g., by accessing the memory via a data bus 505”). Note that: it is known that a computer has at least one processor to execute executable instructions.
Claim 14 is corresponding to the method of claim 6. Therefore, claim 14 is rejected with the same prior art and citations for claim 6.
Regarding claim 18, Bashkirov teaches The one or more non-transitory computer-readable media of claim 11, wherein the character comprises either a virtual character (page 1, para. [0004], “The computer uses the model to compute the exact position and orientation of that certain character, which is eventually rendered into an image. Thus by changing the animation variable values over time, the animator creates motion by making the character move from frame to frame”). Note that: the computer computes pose of the character to render an image as a virtual character in the computer-based environment. or a physical robot. (page 5, para. [0041], “the animation program may use an output of the pose optimization process at 408 to generate robot control inputs 413, as indicated at 414. The animation program may supply the control inputs 413 to a robot controller 415, which converts the control inputs to control signals that are transmitted to an articulated robot 417”; FIG. 4A: “Robot Controller 415” generates the second action of an articulated robot “417”). Note that: the robot controller can control the physical robot in the physical world using the control inputs from the animation program to perform the action.
Regarding claim 19, Bashkirov teaches The one or more non-transitory computer-readable media of claim 11, wherein the environment is at least one of a simulation environment (page 1, para. [0004], “The computer uses the model to compute the exact position and orientation of that certain character, which is eventually rendered into an image. Thus by changing the animation variable values over time, the animator creates motion by making the character move from frame to frame”). Note that: the computer computes pose of the character to render an image as a virtual character in the computer-based simulation environment., an extended reality (XR) environment, a game environment, or a physical environment. (page 5, para. [0041], “the animation program may use an output of the pose optimization process at 408 to generate robot control inputs 413, as indicated at 414. The animation program may supply the control inputs 413 to a robot controller 415, which converts the control inputs to control signals that are transmitted to an articulated robot 417”; FIG. 4A: “Robot Controller 415” generates the second action of an articulated robot “417”). Note that: the robot controller can control the physical robot in the physical world / environment using the control inputs from the animation program to perform the action.
Claim 20 reciting “A system, comprising: one or more memories storing instructions; and one or more processors that are coupled to the one or more memories and, when executing the instructions, are configured to:” is corresponding to the method of claim 1. Therefore, claim 11 is rejected with the same prior art and citations for claim 1.
In addition, Bashkirov teaches A system, comprising: one or more memories storing instructions; and one or more processors that are coupled to the one or more memories and, when executing the instructions, are configured to: (page 6, paras. [0052]-[0053], “The computing device 500 may include one or more processor units 503, which may be configured according to well-known architectures, such as, e.g., single-core, dual-core, quad-core, multi-core, processor-coprocessor, cell processor, and the like. The computing device may also include one or more memory units 504 (e.g., random access memory (RAM), dynamic random access memory (DRAM), read-only memory (ROM), and the like). The processor unit 503 may execute one or more programs 517, portions of which may be stored in the memory 504 and the processor 503 may be operatively coupled to the memory, e.g., by accessing the memory via a data bus 505”). Note that: the device is a system.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 2-5, 7, 9-10, 12, and 15-17 are rejected under 35 U.S.C. 103 as being unpatentable over Bashkirov in view of Yao et al. (ControlVAE: Model-Based Learning of Generative Controllers for Physics-Based Characters, arXiv.org, arXiv:2210.06063v1 [cs.GR] 12 Oct 2022. Hereinafter “Yao”).
Regarding claim 2, Bashkirov discloses The computer-implemented method of claim 1, wherein generating the first action comprises:
encoding the one or more goals to generate one or more tokens; (page 6, para. [0049], “The goal orientations may be represented in axis angle form and encoded into a latent representation using two encoding layers 426, 428”). Note that: the representation can be regarded as tokens to encode the goal orientations (one or more goals) as the inputs.
… the one or more tokens, and one or more masks associated with the one or more tokens … (page 7, claim 12, “wherein an output of the policy network specifies a set of joint position derivatives for one or more joints of the robot”; page 5, para. [0043], “the pose optimization process 408 may use Neural Networks 420 to fit candidate poses in pose sequences 407 to corresponding segmentation masks 403 and, optionally, generate the control inputs 413. In the illustrated implementation, the inputs to the Neural Networks 420 are the segmentation masks 403 obtained from the video frame sequence 401”). Note that: the latent representation as tokens above and the corresponding segmentation masks 403 associated with the latent representation can be combined to formulate feature vectors specifying the character’s features.
However, Bashkirov fails to disclose, but in the same art of computer graphics, Yao discloses
sampling a prior distribution based on a state of the character, the one or more tokens, and one or more masks associated with the one or more tokens to generate a latent vector; and (Yao, page 183:2, col. left, para. 4, “We thus model our VAE-based generative policy with a prior distribution conditional on the simulated character state”; page 183:3, col. right, para. 2, “Figure 2 illustrates the overall structure of ControlVAE. In ControlVAE, a motion control policy is formulated as a conditional distribution 𝜋 (𝒂|𝒔, 𝒛). It computes an action 𝒂 to control the simulated character according to its current state 𝒔 and a latent variable 𝒛 ∈ Z sampled from a prior distribution 𝒛 ∼ 𝑝(𝒛). We train this control policy to allow each latent variable 𝒛 to encode a specific skill, which enables a high-level policy to operate in the latent space Z to accomplish downstream tasks”; page 183:4, Fig. 2: the current state of the character “
PNG
media_image1.png
28
22
media_image1.png
Greyscale
”, prior distribution “
PNG
media_image2.png
42
62
media_image2.png
Greyscale
”, and latent code “
PNG
media_image3.png
24
24
media_image3.png
Greyscale
”, and “
PNG
media_image4.png
434
946
media_image4.png
Greyscale
”; page 183:7, col. left, para. 6, “To encourage the policy to embed diverse motion transitions into prior distribution 𝑝(𝒛|𝒔), we further augment the trajectory collection process with states derived from the prior distribution. Specifically, the policy randomly chooses to sample latent codes from the posterior distribution 𝑞(𝒛𝑡 |𝒔𝑡 , ˜𝒔𝑡+1) or the prior distribution 𝑝(𝒛𝑡 |𝒔𝑡 )”). Note that: (1) the current state of the character
PNG
media_image1.png
28
22
media_image1.png
Greyscale
used in the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
; (2) the token and the corresponding segmentation mask as a features vector above can be regarded as a latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
that is shown in the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
; and (3) sampling the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
can generate the latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
.
processing the latent vector and the state of the character using a decoder included in the trained machine learning model to generate the first action. (Yao, page 183:4, Fig. 2: “Control Policy” as a policy network generate that takes the state of the character (
PNG
media_image1.png
28
22
media_image1.png
Greyscale
) and the latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
as the inputs, and export action “
PNG
media_image5.png
20
24
media_image5.png
Greyscale
” as the first action). Note that: (1) the “Control Policy can be regarded as the trained machine learning model; and (2) “Control Policy” processes the latent vector and the state of the character, and can be regarded as decoder.
Bashkirov and Yao are in the same field of endeavor, namely computer graphics. Before the effective filing date of the claimed invention, it would have been obvious sampling the prior distribution to formulate a latent vector and outputting an action, as taught by Yao into Bashkirov. The motivation would have been “We also learn a state-conditional prior distribution in the VAE-based generative control policy, which generates a skill embedding that outperforms the non-conditional priors in downstream tasks.” (Yao, page 183:1, col. left, para. 1). The suggestion for doing so would allow to sample the prior distribution to formulate a latent vector and output an action. Therefore, it would have been obvious to combine Bashkirov and Yao.
Regarding claim 3, Bashkirov in view of Yao discloses The computer-implemented method of claim 2, wherein sampling the prior distribution comprises:
processing the state of the character, the one or more tokens, and the one or more masks using a prior included in the trained machine learning model to generate a latent distribution; and
sampling the latent vector from the latent distribution. (Yao, page 183:2, col. left, para. 4, “We thus model our VAE-based generative policy with a prior distribution conditional on the simulated character state.”; page 183:3, col. right, para. 2, “Figure 2 illustrates the overall structure of ControlVAE. In ControlVAE, a motion control policy is formulated as a conditional distribution 𝜋 (𝒂|𝒔, 𝒛). It computes an action 𝒂 to control the simulated character according to its current state 𝒔 and a latent variable 𝒛 ∈ Z sampled from a prior distribution 𝒛 ∼ 𝑝(𝒛). We train this control policy to allow each latent variable 𝒛 to encode a specific skill, which enables a high-level policy to operate in the latent space Z to accomplish downstream tasks”; page 183:4, Fig. 2: the current state of the character “
PNG
media_image1.png
28
22
media_image1.png
Greyscale
”, prior distribution “
PNG
media_image2.png
42
62
media_image2.png
Greyscale
”, and latent code “
PNG
media_image3.png
24
24
media_image3.png
Greyscale
”, and “
PNG
media_image4.png
434
946
media_image4.png
Greyscale
”; page 183:4, Fig. 2: “Control Policy” as a policy network generate that takes the state of the character (
PNG
media_image1.png
28
22
media_image1.png
Greyscale
) and the latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
as the inputs, and export action “
PNG
media_image5.png
20
24
media_image5.png
Greyscale
” as the first action). Note that: (1) the current state of the character
PNG
media_image1.png
28
22
media_image1.png
Greyscale
used in the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
; (2) the token code and the corresponding segmentation mask as a features vector above can be regarded as a latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
that is shown in the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
; and (3) sampling the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
can generate the latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
.
The motivation to combine Bashkirov and Yao given in claim 2 is incorporated here.
Regarding claim 4, Bashkirov in view of Yao discloses The computer-implemented method of claim 2, further comprising training a first machine learning model to obtain the trained machine learning model, wherein the first machine learning model comprises an encoder. (Bashkirov, page 1, para. [0006], “techniques have been developed that use motion capture data as a reference in Reinforcement Learning (RL) to train a neural network to control a humanoid robot or create lifelike animations at lower cost”; para. page 6, para. [0049], “The goal orientations may be represented in axis angle form and encoded into a latent representation using two encoding layers 426, 428”). Note that: (1) the representation can be regarded as tokens to encode the goal orientations (one or more goals) as the inputs; (2) two encoding layers 426, 428 can be regarded as an encoder and a first machine learning model; and (3) the first machine model can be trained to obtained the trained machine learning model.
Regarding claim 5, Bashkirov in view of Yao discloses The computer-implemented method of claim 2, wherein sampling the prior distribution comprises sampling random noise and performing one or more reparameterization operations on the random noise to generate the latent vector. (Bashkirov, page 6, para. [0050], “Exploration occurs during training by sampling the policy network output from the learned Gaussian distributions. Sampling in this manner introduces jitter during training that makes learning difficult as it induces falling. The integration scheme discussed above helps to alleviate the jitter. In addition, instead of sampling random actions from the Gaussian distribution at each time step, with fixed probability
PNG
media_image6.png
16
12
media_image6.png
Greyscale
a random action may be sampled from the policy network 422 and with probability 1-
PNG
media_image6.png
16
12
media_image6.png
Greyscale
the robot 417 executes a deterministic action specified by the mean of the Gaussian. Furthermore, updates may be performed using only samples where exploration noise is applied). Note that: (1) although the prior distribution sampling by Bashkirov above is for sampling the policy network output, the same method in sampling the prior distribution to generate the latent vector
PNG
media_image3.png
24
24
media_image3.png
Greyscale
can be applied to the prior distribution
PNG
media_image2.png
42
62
media_image2.png
Greyscale
; (2) a Gaussian distribution is learned and can be used as a prior distribution; and (3) it is obvious to ne having ordinary skill that a random noise corresponding can be sampled on a learned Gaussian distribution that are specified or re-parametrized by a mean of the Gaussian and a corresponding standard deviation.
Regarding claim 7, Bashkirov in view of Yao discloses The computer-implemented method of claim 1, wherein the trained machine learning model comprises at least one of a trained variational autoencoder (VAE) or a trained generative model. (Yao, page 183:1, Fig. 1: “A simulated character gets up, walks, hops, jumps, and runs in our system. We demonstrate a model-based framework for learning a generative motion control policy based on variational autoencoders (VAE), which allows a simulated character to learn a diverse set of skills and use them to accomplish various downstream tasks”). Note that: the trained machine learning model can include a trained VAE after the ControlVAE has been trained for the character’s actions; and (2) ControlVAE is also a generative model.
Regarding claim 9, Bashkirov in view of Yao discloses The computer-implemented method of claim 1, further comprising
wherein the second machine learning model is trained using reinforcement learning to reproduce one or more motions in a set of motion recordings. (Bashkirov, FIG 4B: typical “Policy Network” and “Critic Network” forrms a second machine leaning model using reinforcement learning to reproduce one or more motions in a set of motion recordings; page 4, para. [0037], “FIG. 4A depicts an example of a generalized method for monocular pose prediction in computer animation according to aspects of the present disclosure. The method may begin with an input video sequence of frames 401. The input video frame may be obtained from a live feed or from archival footage. Any suitable type of video frame that shows a character may be used. Preferably, the input video frame sequence 401 is in the form of frames of digital video. Alternatively, a non-digital video frame or motion picture frame may be digitized to provide the input video frame sequence 401. An animation program may generate a corresponding sequence of segmentation masks 403 of a character in each frame of the input video frame sequence 401, as indicated at 402. The segmentation mask 403 may be an edge mask”). Note that: (1) the motions in the frames of video recording can be used to train the second machine learning model to reproduce
… training a first machine learning model to produce the trained machine learning model based on a loss that is a metric of comparison between actions generated by the first machine learning model and actions generated by a second machine learning model, (Yao, page 18:3, Fig. 2, “controlVAE“ can be a first machine learning model; page 183:5,
PNG
media_image7.png
758
598
media_image7.png
Greyscale
”). Note that: the first machine learning model (ControlVAE System) is trained using the loss function indicated by equation (13) in which
PNG
media_image8.png
28
42
media_image8.png
Greyscale
indicated by equation (10) above is a metric of comparison between actions generated by the first machine learning model (ControlVAE) and actions as reference (ground truth) generated by a second machine learning model while the second machine learning model after its training has been done can be used to generate the reference actions.
The motivation to combine Bashkirov and Yao given in claim 2 is incorporated here.
Regarding claim 10, Bashkirov in view of Yao discloses The computer-implemented method of claim 9, wherein the first machine learning model is trained using one or more motions that are sampled from a set of motion recordings, and the one or more motions are masked based on one or more sampled masks. (Bashkirov, page 4, para. [0037], “FIG. 4A depicts an example of a generalized method for monocular pose prediction in computer animation according to aspects of the present disclosure. The method may begin with an input video sequence of frames 401. The input video frame may be obtained from a live feed or from archival footage. Any suitable type of video frame that shows a character may be used. Preferably, the input video frame sequence 401 is in the form of frames of digital video. Alternatively, a non-digital video frame or motion picture frame may be digitized to provide the input video frame sequence 401. An animation program may generate a corresponding sequence of segmentation masks 403 of a character in each frame of the input video frame sequence 401, as indicated at 402. The segmentation mask 403 may be an edge mask”). Note that: (1) the first machine learning model can be trained using the frame data showing the motions extracted from video recordings of archival footage; and (2) a corresponding sequence of segmentation masks 403 mask the one or more motions.
Claim 12 is corresponding to the method of claim 2. Therefore, claim 12 is rejected for the same rationale for claim 2, respectively.
Claims 15-16 are corresponding to the method of claims 9-10, respectively. Therefore, claims 15-16 are rejected for the same rationale for claims 9-10, respectively.
Regarding claim 17, Bashkirov in view of Yao discloses The one or more non-transitory computer-readable media of claim 15, wherein the training the first machine learning model further comprises increasing a value of a Kullback-Leibler (KL)-coefficient during successive iterations of the training. (Yao, page 183:5, col. left, para. 5, “
PNG
media_image9.png
72
544
media_image9.png
Greyscale
; page 183:5, col. right, para. 1, “Following 𝛽-VAE [Higgins et al. 2017], we employ a weight parameter 𝛽 for the KL-divergence loss, which increases once every 500 training epochs from 0.01 to 0.1 during the training.”). Note that: during training the first machine learning model, a weight parameter 𝛽 as a value of a Kullback-Leibler (KL)-coefficient for the KL-divergence loss increases during successive iterations of the training.
The motivation to combine Bashkirov and Yao give in claim 12 is incorporated here.
Claims 13 are rejected under 35 U.S.C. 103 as being unpatentable over Bashkirov in view of Yao, Lacoste et al. (Deep Prior, arXiv.org, arXiv:1712.05016v2 [stat.ML] 16 Dec 2017, hereinafter “Lacoste”), and Nvidia (What Is a Transformer Model, Archive.org, https://web.archive.org/web/20240318182121/https://blogs.nvidia.com/blog/what-is-a-transformer-model/, hereinafter “Nvidia”) .
Regarding claim 13, Bashkirov in view of Yao discloses The one or more non-transitory computer-readable media of claim 12, wherein
the decoder comprises a fully-connected neural network. (Yao, “Control Policy” is the decoder above; Fig. 2, “
PNG
media_image4.png
434
946
media_image4.png
Greyscale
”). Note that: (1) as shown in Fig. 2 above for the decoder (Control Policy), it is a fully connected neural network as it really is; and (2) it is obvious to one having ordinary skill in the art, a neural network like “Control Policy” can be implemented using either a fully connected neural network or other neural networks.
The motivation to combine Bashkirov and Yao given in claim 2 is incorporated here.
However, Bashkirov in view of Yao fails to disclose, but in the same art of neural network application, Lacoste discloses
the prior distribution is generated by a prior that comprises a (Lacoste, page 1, Abstract, “In this work we investigate the possibility of learning the prior distribution over neural network parameters using such tools”). Note that: a neural network can be trained to learn or generate a prior distribution, and the neural network can be regarded as a prior.
Bashkirov in view of Yao, and Lacoste, are in the same field of endeavor, namely neural network application. Before the effective filing date of the claimed invention, it would have been obvious learning a prior distribution using a neural network, as taught by Lacoste into Bashkirov in view of Yao. The motivation would have been “In this work we investigate the possibility of learning the prior distribution over neural network parameters using such tools” (Lacoste, page 1, Abstract). The suggestion for doing so would allow to generate a prior distribution using a neural network as a prior. Therefore, it would have been obvious to combine Bashkirov, Yao, and Lacoste.
However, Bashkirov in view of Yao and Lacoste fails to disclose, but in the same art of neural network application, Nvdia discloses transformer-based neural network (Nvdia, page 1, para. 4, “A transformer model is a neural network that learns context and thus meaning by tracking relationships in sequential data like the words in this sentence.”; page 3, para. 4, “Transformers are in many cases replacing convolutional and recurrent neural networks (CNNs and RNNs), the most popular types of deep learning models just five years ago.”). Note that: Transformer-based neural network can be used to replace convolutional and recurrent neural networks to learn context and thus meaning by tracking relationships.
Bashkirov in view of Yao and Lacoste, and Nvidia, are in the same field of endeavor, namely neural network application. Before the effective filing date of the claimed invention, it would have been obvious using Transformer-based neural network to replace convolutional and recurrent neural networks, as taught by Nvdia into Bashkirov in view of Yao and Lacoste. The motivation would have been “Transformers are in many cases replacing convolutional and recurrent neural networks (CNNs and RNNs), the most popular types of deep learning models just five years ago.” (Nvidia, page 3, para. 4). The suggestion for doing so would allow to use Transformer-based neural network to replace convolutional and recurrent neural networks. Therefore, it would have been obvious to combine Bashkirov, Yao, Lacoste, and Nvidia.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Jiang et al. (VIMA: General Robot Manipulation with Multimodal Prompts, ArXiv.org, arXiv:2210.03094v2 [cs.RO] 28 May 2023) teaches a transformer-based robot agent, VIMA, that processes these prompts and outputs motor actions autoregressively.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BIAO CHEN whose telephone number is (703)756-1199. The examiner can normally be reached M-F 8am-5pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kee M Tung can be reached at (571)272-7794. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Biao Chen/
Patent Examiner, Art Unit 2611
/KEE M TUNG/Supervisory Patent Examiner, Art Unit 2611
Prosecution Insights