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 .
Status of the Claims
Claims 1-9, 11-24 are currently pending in the present application, with claims 1, 21, 22, and 24 being independent.
Response to Amendments / Arguments
Applicant’s arguments, see Pg. 10, filed 11/26/2025, with respect to the abstract have been fully considered and are persuasive. The objection to the abstract has been withdrawn.
Applicant’s arguments, see Pg. 11, filed 11/26/2025, with respect to claim 10 have been fully considered and are persuasive. The objection of claim 10 has been withdrawn
Applicant’s arguments, see Pg. 11, filed 11/26/2025, with respect to 35. U.S.C. § 112f interpretation have been fully considered and are persuasive. Examiner agrees that structural support for such modules may be found in Fig. 1, 3-5, 8, and 10-11, and associated text in [0105]-[0111] of the present applicant.
Applicant’s arguments, see Pg. 11-12, filed 11/26/2025, with respect to claims 1-24 have been fully considered and are persuasive. The rejection under 35 U.S.C. §101 of claims 1-24 has been withdrawn
Applicant's arguments, see Pg. 12-18, filed 11/26/2025 have been fully considered but they are not persuasive.
Applicant argues: Wang et al. "Humanise: Language-conditioned human motion generation in 3d scenes." Advances in Neural Information Processing Systems 35 (2022): 14959-14971, does not disclose “a prediction module configured to generate predicted trajectories of the human performing the action based on the encodings using a latent space” and further applicant submits that Wang’s Gaussian distribution parameters are not equivalent to the claimed “trajectories of the human performing the action based on the encoding”.
Examiner replies: The claimed “prediction module” reasonably reads on Wang’s predicted time-sequence of motion/body parameters output for the human motions. In Section 4.2 and Fig. 3 of Wang:
Wang’s motion encoder “obtains a sequence-level feature of the input motionΘ1: T …concatenated with the conditional embedding z_c, followed by an MLP layer to predict the Gaussian distribution parameters (i.e., μ and Σ)…” in Pg. 6, Section Motion encoder.
Wang’s motion decoder “use a transformer decoder to generate a sequence of parameters Θ1:T for a given duration T… “use T sinusoidal positional embeddings to query the concatenation of the sampled latent z and the conditional embedding z_c”, which is used to generate a time-ordered sequence of motion outputs.
Under broadest reasonable interpretation, the “encodings using a latent space” is the learned latent variable space represented by sampling z ~ N(μ,Σ) as shown in Fig. 3. The “predicted trajectories” are the predicted time series motion outputs (the sequence of poses/SMPL-X parameters) generated by the decoder, and those trajectories are “based on the encodings” because the encoder output drives μ,Σand therefore z, which conditions the trajectory output. The prediction is realized by the combination of latent inference and decoding. Further, the claim does not require a physically separate “module” distinct from the encoder, it only requires functionality, that Wang’s motion encoder and motion decoder collectively perform the claimed “prediction” module. Therefore, Wang’s architecture of the latent inference (μ,Σ, z) and decoder stage that produces the predicted sequence reasonably corresponds to the “prediction module” as claimed.
Applicant argues: Wang does not disclose “a decoder module configured to generate decodings based on the predicted trajectories; and a rendering module configured to generate the rendering based on the decodings”.
Examiner replies: Wang expressly discloses the the transformer decoder outputs a time sequence of predicted body parameters, and those predicted parameters are then mapped into body meshes using differentiable SMPL-X model. Specifically:
“use a transformer decoder to generate a sequence of parameters bΘ1:T for a given duration T…The outputs of the transformer decoder are mapped into body meshes with differentiable SMPL-X model” in Wang Pg. 6, Section 4.2, Motion decoder and Fig. 3.
Under broadest reasonable interpretation, the body meshes produced from the predicted parameter sequence are “decodings,” and generating a visual output from those meshes reads on “generating the rendering based on the decodings”. Therefore, Wang’s architecture produces mesh outputs from predicted motion parameters specifically for producing the visualized human motion results (Pg. 10, Section 5.4; the start and the end human body poses as input; the goal is to predict all human poses in between…outputs the parameters for all the human poses along the route), which satisfies the functional requirement of “a decoder module configured to generate decodings based on the predicted trajectories; and a rendering module configured to generate the rendering based on the decodings”.
Applicant argues: Wang does not disclose “(a) train the model based on input video including humans performing actions”.
Examiner replies: Wang expressly discloses that its learning is based on human motion sequences and describes its dataset and statistics in terms of motion sequences and frames: “HUMANISE includes 19.6k human motion sequences in 643 3D scenes” on Pg. 5, Section 3.3 and Table 1.
Wang also expresses the model input is explicitly: “the input motion Θ1:T” on Pg. 6, Section 4.2 Motion encoder.
Further, Wang’s disclosure does not merely state a loss abstractly, but additionally describes the training procedure “We train our generative model on HUMANISE for 150 epochs...training the action-agnostic model on the entire HUMANISE dataset” on Pg. 7, Section 4.4. Wang also demonstrates the trained model in experiments using these human motion sequences “To evaluate the model performance, we conduct experiments in both the action-specific and action-agnostic settings. In the action-specific setting, we train and evaluate our model on HUMANISE subsets of each action” on Pg. 7, Section 5.1, which necessarily discloses that the model was trained using the disclosed HUMANISE motion sequences and training setup (Wang Section 4.4 and Section 5). Accordingly, Wang teaches training the system based on input video including humans performing actions.
Applicant argues: Wang does not disclose “after (a), train the model based on geometry of a scene, one or more target actions for performance by a human in the scene, and observations of the human during performance of the one or more target actions in the scene”.
Examiner replies: Wang expressly discloses incorporating scene geometry and action/interaction semantics into the learning pipeline through its condition module: “The condition module takes input from two modalities (i.e., the given scene S and the language description L)” and processes the 3D scene using a “Point Transformer” on Pg. 5, Section 4.2, thereby, using the scene geometry representation as an input driving the learned conditioning embedding.
Further, Wang discloses training losses directed to action/object grounding, stating “We devise the training loss with three components: reconstruction loss, KL divergence loss, and two auxiliary losses for object grounding and action-specific generation” and “we introduce two auxiliary tasks for locating the interacting target object and generating action-specific motions” on Pg. 6, Section 4.3. And Wang’s explicit reciting of a training procedure “We train our generative model on HUMANISE for 150 epochs...training the action-agnostic model on the entire HUMANISE dataset” on Pg. 7, Section 4.4. Wang also demonstrates the trained model in experiments using these human motion sequences “To evaluate the model performance, we conduct experiments in both the action-specific and action-agnostic settings. In the action-specific setting, we train and evaluate our model on HUMANISE subsets of each action” on Pg. 7, Section 5.1 provide explicit confirmation that the model is trained and evaluated in this scene-conditioned, action-conditioned setting, that necessarily incorporates the scene representation S (Wang Section 4.4; Our point transformer module that processes the scene point cloud…we pre-train the point transformer with a semantic segmentation task on ScanNet) and the action/interaction description (Wang Section 4.4; We employ the official pre-trained BERT model to process the language descriptions). Accordingly, Wang teaches the limitations of “after (a), train the model based on geometry of a scene, one or more target actions for performance by a human in the scene, and observations of the human during performance of the one or more target actions in the scene” (Wang Sections 3.1, 4.2-4.4, and 5). To the extent Applicant respectfully asserts the term “after (a)” requires a distinct sequential phase 2, the claim language does not require separate training runs, but merely recites that the training module is configured to train based on (a) and then based on (b), which is satisfied where the training incorporate the input motion (Pg. 6, Section 4.2 Motion encoder) together with the inputs from the condition modules (Pg. 5, Section 4.2 Condition module), and further expressly discloses observations and evaluation metrics of human action-specific and action-agnostic settings (Wang Section 5.1-5.2 and Fig. 4).
Applicant argues: Wang does not disclose “(a) training a latent space using video including humans performing actions; and (b), after (a), using the trained latent space, training a model configured to generate a rendering of a human performing an action in a space using based on geometry of a scene, one or more target actions for performance by a human in the scene, and observations of the human during performance of the one or more target actions in the scene”.
Examiner replies: Wang’s encoder expressly predicts “Gaussian distribution parameters μ andΣ” from the motion encoding and then “sample a latent vector z from the distribution using the reparameterization trick…” (Wang Section 4.2, Pg. 6), which explicitly expresses learning and using latent space representation. Wang also explicitly identifies and uses human motion sequences and frames as the training data “19.6k human motion sequences” (Wang, Section 3.3 and Table 1) and “The condition module takes input from two modalities (i.e., the given scene S and the language description L)” and processes the 3D scene using a “Point Transformer” on Pg. 5, Section 4.2, thereby, using the scene geometry representation as an input driving the learned conditioning embedding. Wang further discloses the training and evaluation procedure for the disclosed architecture and dataset in Sections 4.4 and Section 5.
Regarding the remaining arguments: Applicant argues with respect to the amended claim language, which is fully addressed in the prior art rejections set forth below.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The following is a quotation of pre-AIA 35 U.S.C. 112, sixth paragraph:
An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph:
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action.
This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, because the claim limitation(s) uses a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier. Such claim limitation(s) is/are:
In claim 1:
“a model configured to generate”
“an encoder module configured to encode”
“a prediction module configured to generate”
“a decoder module configured to generate”
“a rendering module configured to generate”
“a training module configured to train”
In claims 2, 4, 8-10, 12:
“the training module configured to train”
In claim 5, 14, 16, 19-20:
“the training module is further configured to train”
In claim 13:
“the training module is configured to determine”
In claim 18:
“the encoder module is configured to add”
In claim 21:
“a model configured to generate”
In claim 22:
“a generator module configured to encode”
“a prediction module configured to generate”
“a rendering module configured”
In claim 24:
“the model configured to generate”
“an encoder module configured to encode I”
“a prediction module configured to generate”
“a decoder configured to generate”
“a rendering module configured to generate”
Because this/these claim limitation(s) is/are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, it/they is/are being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof (see Par. 0064 on model, Par. 0056 and 0087 on encoder module, Par. 0065, 0081, and 0088 on prediction module, Par. 0069, 0072 and 0086 on decoder module, Par. 0058 and 0064 on rendering module, Par. 0060-0061, 0067-0068, and 0095 on training module, and Par. 0097-0098 on generator module). Further, as recited in examiner’s response to arguments, additional structural support for such modules may be found in Fig. 1, 3-5, 8, and 10-11, and associated text in [0105]-[0111] of the originally filed disclosure.
If applicant does not intend to have this/these limitation(s) interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, applicant may: (1) amend the claim limitation(s) to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitation(s) recite(s) sufficient structure to perform the claimed function so as to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph.
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.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1-2, 6, 8-9, 11-24 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Wang et al. “HUMANISE: Language-conditioned human motion generation in 3d scenes”. In NeurIPS, 2022, hereinafter referred to as “Wang”.
Regarding claim 1, Wang discloses a motion generation system (Neural Information Processing Systems), comprising:
a model configured to generate a rendering of a human performing an action in a space (Fig. 3 and Pg. 2, Section 1; HUMANISE…language-conditioned human motion generation in 3D scenes) the model including:
an encoder module configured to encode input into encodings (Pg. 2, Section 1; the input motion is encoded with a recurrent module. Pg. 6, Section 4.2; Motion encoder),
a prediction module configured to generate predicted trajectories of the human performing the action based on the encodings using a latent space (Pg. 6, Section 4.2; the output is concatenated with the conditional embedding z_c, followed by an MLP layer to predict the Gaussian distribution parameters (i.e., µ and Σ). Finally, we sample a latent vector z from the distribution using the reparameterization trick),
Examiner’s note: Additionally, as recited above in examiner’s response to arguments, the claimed “prediction module” reasonably reads on Wang’s predicted time-sequence of motion/body parameters output for the human motions. In Section 4.2 and Fig. 3 of Wang: Wang’s motion encoder “obtains a sequence-level feature of the input motionΘ1: T …concatenated with the conditional embedding z_c, followed by an MLP layer to predict the Gaussian distribution parameters (i.e., μ and Σ)…” in Pg. 6, Section Motion encoder. Wang’s motion decoder “use a transformer decoder to generate a sequence of parameters Θ1:T for a given duration T… “use T sinusoidal positional embeddings to query the concatenation of the sampled latent z and the conditional embedding z_c”, which is used to generate a time-ordered sequence of motion outputs. Under broadest reasonable interpretation, the “encodings using a latent space” is the learned latent variable space represented by sampling z ~ N(μ,Σ) as shown in Fig. 3. The “predicted trajectories” are the predicted time series motion outputs (the sequence of poses/SMPL-X parameters) generated by the decoder, and those trajectories are “based on the encodings” because the encoder output drives μ,Σand therefore z, which conditions the trajectory output. The prediction is realized by the combination of latent inference and decoding. Further, the claim does not require a physically separate “module” distinct from the encoder, it only requires functionality, that Wang’s motion encoder and motion decoder collectively perform the claimed “prediction” module. Therefore, Wang’s architecture of the latent inference (μ,Σ, z) and decoder stage that produces the predicted sequence reasonably corresponds to the “prediction module” as claimed.
a decoder module configured to generate decodings based on the predicted trajectories (Pg. 2, Section 1; and decoded with a transformer decoder to generate human motions. Pg. 6, Section 4.2; Motion decoder),
and a rendering module configured to generate the rendering based on the decodings (Pg. 6, Section 4.2; the outputs of the transformer decoder are mapped into body meshes with the differentiable SMPL-X model),
Examiner’s note: Additionally, as recited above in examiner’s response to arguments, the “use a transformer decoder to generate a sequence of parameters bΘ1:T for a given duration T…The outputs of the transformer decoder are mapped into body meshes with differentiable SMPL-X model” in Wang Pg. 6, Section 4.2, Motion decoder and Fig. 3 under broadest reasonable interpretation, the body meshes produced from the predicted parameter sequence are “decodings,” and generating a visual output from those meshes reads on “generating the rendering based on the decodings”. Therefore, Wang’s architecture produces mesh outputs from predicted motion parameters specifically for producing the visualized human motion results (Pg. 10, Section 5.4; the start and the end human body poses as input; the goal is to predict all human poses in between…outputs the parameters for all the human poses along the route), which satisfies the functional requirement of “a decoder module configured to generate decodings based on the predicted trajectories; and a rendering module configured to generate the rendering based on the decodings”.
one or more processors; and
memory including instructions that, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU):
(a) train the model based on input video including humans performing actions (Pg. 3, Section 3; we propose a synthetic dataset by leveraging the existing datasets in human motion (i.e., AMASS [Mahmood et al. 2019]). Pg. 6, Section 4.2; input motion Θ_1:T),
Examiner’s note: Additionally, as recited above in examiner’s response to arguments, Wang expressly discloses that its learning is based on human motion sequences and describes its dataset and statistics in terms of motion sequences and frames: “HUMANISE includes 19.6k human motion sequences in 643 3D scenes” on Pg. 5, Section 3.3 and Table 1. Wang also expresses the model input is explicitly: “the input motion Θ1:T” on Pg. 6, Section 4.2 Motion encoder. Further, Wang’s disclosure does not merely state a loss abstractly, but additionally describes the training procedure “We train our generative model on HUMANISE for 150 epochs...training the action-agnostic model on the entire HUMANISE dataset” on Pg. 7, Section 4.4. Wang also demonstrates the trained model in experiments using these human motion sequences “To evaluate the model performance, we conduct experiments in both the action-specific and action-agnostic settings. In the action-specific setting, we train and evaluate our model on HUMANISE subsets of each action” on Pg. 7, Section 5.1, which necessarily discloses that the model was trained using the disclosed HUMANISE motion sequences and training setup (Wang Section 4.4 and Section 5). Accordingly, Wang teaches training the system based on input video including humans performing actions.
and (b), after (a), train the model based on geometry of a scene (Pg. 5, Section 4.2; given scene S…to process the 3D scene, we employ the Point Transformer to produce point-level features. Pg. 7, Section 4.4; Point transformer), one or more target actions for performance by a human in the scene (Pg. 5, Section 4.2; language description L…we use the pre-trained BERT to extract the word-level embedding. Pg. 7, Section 4.4; BERT), and observations of the human during performance of the one or more target actions in the scene (Fig. 2, 4, 6 and Pg. 5, Section 4.2; condition module takes input from two modalities (i.e., the given scene S and the language description L)…features are concatenated and fed into the feature fusion module…mapped to a conditional latent embedding z_c).
Examiner’s note: Additionally, as recited above in examiner’s response to arguments Wang expressly discloses incorporating scene geometry and action/interaction semantics into the learning pipeline through its condition module: “The condition module takes input from two modalities (i.e., the given scene S and the language description L)” and processes the 3D scene using a “Point Transformer” on Pg. 5, Section 4.2, thereby, using the scene geometry representation as an input driving the learned conditioning embedding.
Further, Wang discloses training losses directed to action/object grounding, stating “We devise the training loss with three components: reconstruction loss, KL divergence loss, and two auxiliary losses for object grounding and action-specific generation” and “we introduce two auxiliary tasks for locating the interacting target object and generating action-specific motions” on Pg. 6, Section 4.3. And Wang’s explicit reciting of a training procedure “We train our generative model on HUMANISE for 150 epochs...training the action-agnostic model on the entire HUMANISE dataset” on Pg. 7, Section 4.4. Wang also demonstrates the trained model in experiments using these human motion sequences “To evaluate the model performance, we conduct experiments in both the action-specific and action-agnostic settings. In the action-specific setting, we train and evaluate our model on HUMANISE subsets of each action” on Pg. 7, Section 5.1 provide explicit confirmation that the model is trained and evaluated in this scene-conditioned, action-conditioned setting, that necessarily incorporates the scene representation S (Wang Section 4.4; Our point transformer module that processes the scene point cloud…we pre-train the point transformer with a semantic segmentation task on ScanNet) and the action/interaction description (Wang Section 4.4; We employ the official pre-trained BERT model to process the language descriptions). Accordingly, Wang teaches the limitations of “after (a), train the model based on geometry of a scene, one or more target actions for performance by a human in the scene, and observations of the human during performance of the one or more target actions in the scene” (Wang Sections 3.1, 4.2-4.4, and 5). To the extent Applicant respectfully asserts the term “after (a)” requires a distinct sequential phase 2, the claim language does not require separate training runs, but merely recites that the training module is configured to train based on (a) and then based on (b), which is satisfied where the training incorporate the input motion (Pg. 6, Section 4.2 Motion encoder) together with the inputs from the condition modules (Pg. 5, Section 4.2 Condition module), and further expressly discloses observations and evaluation metrics of human action-specific and action-agnostic settings (Wang Section 5.1-5.2 and Fig. 4).
Regarding claim 2, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when executed by the one or more processors, causes the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU), during (a), the training module is configured to train the encoder module (Pg. 6, Section 4.2; Motion encoder: we first use a bidirectional GRU to obtain a sequence level feature of the input motion Θ_1:T) and the decoder module (Pg. 6, Section 4.2; Motion decoder: …transformer decoder…mapped into body meshes) based on the input video (Fig. 3).
Regarding claim 6, Wang discloses the motion generation system of claim 1 and further discloses wherein the encoder module includes an auto-regressive encoder (Pg. 6, Section 4.1; Motion encoder: we first use a bidirectional GRU to obtain a sequence level feature. Examiner's note: GRU is the autoregressive recurrent encoder).
Regarding claim 8, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU), train the latent space during (a) based on the input video including humans performing actions (Pg. 6, Section 4.2-4.3; bidirectional GRU to obtain a sequence-level feature of the input motion Θ_1:T …sample a latent vector z from the distribution…latent z).
Regarding claim 9, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when exected by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU), train the encoder module and the decoder module (Fig. 3) during (b) based on the geometry of the scene (Pg. 5, Section 4.2; given scene S…), the one or more target actions for performance by the human in the scene (Pg. 5, Section 4.2; language description L…), and the observations of the human during performance of the one or more target actions in the scene (Fig. 2, 4, 6 and Pg. 5, Section 4.2; condition module takes input from two modalities (i.e., the given scene S and the language description L)…features are concatenated and fed into the feature fusion module…mapped to a conditional latent embedding z_c).
Regarding claim 11, Wang discloses the motion generation system of claim 1 and further discloses wherein the observations initially include a target position and pose of the human (Pg. 4, Section 3.1; final pose) in the scene (Fig. 2 and Pg. 4, Section 3.1; The language description in HUMANISE provides instructions to walk to a target location, usually to an object; e.g. walk to the refrigerator. We sample dense positions near the specified object on the floor as the potential targets for the final pose of walking motion).
Regarding claim 12, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) train the model based on minimizing a prediction loss during (b) (Pg. 6, Section 4.3; Training loss…three components: reconstruction loss, KL divergence loss, and two auxiliary loss for object grounding and action-specific generation).
Regarding claim 13, Wang discloses the motion generation system of claim 12 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) determine the prediction loss based on an output of the encoder and a predetermined output (Pg. 6, Section 4.3; reconstruction loss…Eqn. (2) …distance between the predicted body parameters and the ground truth…).
Regarding claim 14, Wang discloses the motion generation system of claim 12 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) train the model based on minimizing a contact loss during (b) (Pg. 4, Section 3.1; Contact constraint).
Regarding claim 15, Wang discloses the motion generation system of claim 14 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) to minimize the contact loss based on increasing contact between the human and an object in the scene (Pg. 4, Section 3.1; the foot-ground support constraint and the proximity between the human and the interacting object. The support constraint is a distance threshold between the foot and the ground plane. The human-object proximity differs among actions; we design the following action-specific proximity constraints for realistic interaction…).
Regarding claim 16, Wang discloses the motion generation system of claim 12 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) train the model based on minimizing an interpenetration loss during (b) (Pg. 4, Section 3.1; Collision constraint).
Regarding claim 17, Wang discloses the motion generation system of claim 16 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) minimize the interpenetration loss based on preventing the human from penetrating an object in the scene (Pg. 4, Section 3.1; Collision avoidance of human motions is equivalent to finding the closest point to the queried human vertex in the scene by searching through the k-d tree. Collision happens as the distance between the human vertex and the closest point in the scene is less than a threshold. A valid motion should avoid collisions with the objects or walls for all the poses along the trajectory…).
Regarding claim 18, Wang discloses the motion generation system of claim 1 and further discloses wherein the encoder module is configured to add time dependent encodings to frames of video (Pg. 6, Section 4.2; transformer decoder…use T sinusoidal positional embeddings to query the concatenation of the sampled latent z and the conditional embedding z_c…duration T. Examiner's note: in the context of transformer architecture, "time" refers to the sequence position, sinusoidal positional embeddings are the time-dependent encodings).
Regarding claim 19, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) during (b) train the model further based on a target path of the human in the scene (Fig. 6 and Pg. 10, Section 5.4; RouteNet takes the scene point cloud, the start and the end locations and orientations as inputs; it generates a possibly route between the start and end location…outputs the parameters or all the human poses along the route).
Regarding claim 20, Wang discloses the motion generation system of claim 1 and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) during (b) train the model further based on one or more future observations of the human during performance of the one or more target actions in the scene (Fig. 4 and Pg. 8, Section 5.2; we use the human body in the first frame to compute the body-to-goal distance for stand up action and use the last frame for other actions).
Regarding claim 21, Wang discloses a non-transitory computer readable medium including instructions for performing a training method for a model configured to generate renderings of humans (Pg. 5, Section 3.3 and Fig. 2; HUMANISE includes 19.6k human motion sequencies in 643 3D scenes. The four most representative actions are sit (5578), stand up (3463), lie down (2343), and walk (8264) …Pg. 7, Section 5.1; we split motions in HUMANISE according to the original scene IDs and split in ScanNet…resulting in 16.5k motions in 543 scenes for training), the training method comprising:
(a) training a latent space using video including humans performing actions (Pg. 6, Section 4.2-4.3; bidirectional GRU to obtain a sequence-level feature of the input motion Θ_1:T…sample a latent vector z from the distribution…reconstruction loss, KL divergence loss) and
(b), after (a), using the trained latent space, training a model configured to generate a rendering of a human performing an action in a space (Pg. 6, Section 4.3; reconstruction loss, KL divergence loss, and two auxiliary losses for object grounding and action-specific generation) using based on geometry of a scene (Pg. 5, Section 4.2; given scene S), one or more target actions for performance by a human in the scene (Pg. 5, Section 4.2; language description L, and observations of the human during performance of the one or more target actions in the scene (Fig. 2, 4, 6 and Pg. 5, Section 4.2; condition module takes input from two modalities (i.e., the given scene S and the language description L)…features are concatenated and fed into the feature fusion module…mapped to a conditional latent embedding z_c).
Examiners note: Additionally, as recited above in examiner’s response to arguments, Wang’s encoder expressly predicts “Gaussian distribution parameters μ andΣ” from the motion encoding and then “sample a latent vector z from the distribution using the reparameterization trick…” (Wang Section 4.2, Pg. 6), which explicitly expresses learning and using latent space representation. Wang also explicitly identifies and uses human motion sequences and frames as the training data “19.6k human motion sequences” (Wang, Section 3.3 and Table 1) and “The condition module takes input from two modalities (i.e., the given scene S and the language description L)” and processes the 3D scene using a “Point Transformer” on Pg. 5, Section 4.2, thereby, using the scene geometry representation as an input driving the learned conditioning embedding. Wang further discloses the training and evaluation procedure for the disclosed architecture and dataset in Sections 4.4 and Section 5.
Regarding claim 22, Wang discloses a motion generation system configured to generate a rendering of a human performing a target action (Fig. 3 and Pg. 2, Section 1; HUMANISE…language-conditioned human motion generation in 3D scenes) in an inference scene (Pg. 6, Section 4.2; duration T equals the length of the input motion sequence in the training phase, it can be arbitrary values during inference), comprising:
one or more processors; and
memory including instructions that, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU):
encode input into encodings (Pg. 2, Section 1; the input motion is encoded with a recurrent module. Pg. 6, Section 4.2; Motion encoder), the input including the target action to be performed in the inference scene and a geometry of the inference scene (Pg. 5, Section 4.2 and Fig. 3; condition module takes input from two modalities (i.e., the given scene S and the language description L…the scene and language features are finally concatenated and mapped to a conditional latent embedding z_c),
generate predicted trajectories of the human performing the target action based on the encodings using a latent space (Pg. 6, Section 4.2; the output is concatenated with the conditional embedding z_c, followed by an MLP layer to predict the Gaussian distribution parameters (i.e., µ and Σ). Finally, we sample a latent vector z from the distribution using the reparameterization trick),
generate decodings based on the predicted trajectories (Pg. 2, Section 1; and decoded with a transformer decoder to generate human motions. Pg. 6, Section 4.2; Motion decoder), and
generate the rendering based on the decodings (Pg. 6, Section 4.2; the outputs of the transformer decoder are mapped into body meshes with the differentiable SMPL-X model),
wherein the motion generation system is trained based on:
(a) input video including humans performing one or more training actions in a training scene (Pg. 3, Section 3; we propose a synthetic dataset by leveraging the existing datasets in human motion (i.e., AMASS [Mahmood et al. 2019]). Pg. 6, Section 4.2; input motion Θ_1:T), and
(b), after training based on (a), based on geometry of the training scene (Pg. 5, Section 4.2; given scene S…to process the 3D scene, we employ the Point Transformer to produce point-level features. Pg. 7, Section 4.4; Point transformer), the one or more training actions for performance by a human in the training scene (Pg. 5, Section 4.2; language description L…we use the pre-trained BERT to extract the word-level embedding. Pg. 7, Section 4.4; BERT), and observations of the human during performance of the one or more training actions in the training scene (Fig. 2, 4, 6 and Pg. 5, Section 4.2; condition module takes input from two modalities (i.e., the given scene S and the language description L) …features are concatenated and fed into the feature fusion module…mapped to a conditional latent embedding z_c).
Regarding claim 23, Wang discloses the motion generation system of claim 22 and further discloses the input further including past observations of the human in the inference scene (Pg. 6, Section 4.2; input motion Θ_1:T. Pg. 6, Section 4.2; duration T equals the length of the input motion sequence in the training phase, it can be arbitrary values during inference).
Regarding claim 24, claim 24 is the CRM claim of system claim 1 and is accordingly rejected using substantially similar rationale as to that which is set for with respect to claim 1.
Claim Rejections - 35 USC § 103
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 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.
Claim(s) 3-5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang et al. “HUMANISE: Language-conditioned human motion generation in 3d scenes”. In NeurIPS, 2022, hereinafter referred to as “Wang”, in view of Razavi et al. "Generating diverse high-fidelity images with vq-vae-2." Advances in neural information processing systems 32 (2019), hereinafter referred to as “Razavi”.
Regarding claim 3, Wang discloses the motion generation system of claim 2, but does not disclose wherein the encoder module includes a quantizer.
In the same art of VAEs, Razavi discloses wherein the encoder module includes a quantizer (Pg. 2, Section 2.1; VQ-VAE model…an encoder that maps observations onto a sequence of discrete variables…the encoder is a non-linear mapping from the input space, x, to a vector E(x). This vector is then quantized based on its distance to the prototype vectors in the codebook…Eqn. (1)).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the quantizer of Razavi into Wang’s VAE framework. Doing so provides a well-known technique to improve training stability and produce more precise, high-quality, and discrete latent representations. By quantizing the continuous latent space into discrete codes, VQ-VAEs create a more scalable system suitable for high-fidelity generation tasks. Incorporating this into Wang’s motion generation system would yield predictable results in providing higher-quality latent features for high-fidelity human motion rendering in 3D scenes.
Regarding claim 4, Wang discloses the motion generation system of claim 2, but does not disclose wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) train the encoder module based on minimizing a difference between a discrete latent sequence and a latent sequence output by the encoder module.
In the same art of VAEs, Razavi discloses wherein the training module is configured to train the encoder module based on minimizing a difference between a discrete latent sequence and a latent sequence output by the encoder module (Pg. 2, Section 2.1; The VQ-VAE model incorporates two additional terms in its objective to align the vector space of the codebook with the output of the encoder. The codebook loss…brings the selected codebook e close to the output of the encoder, E(x). The commitment loss…encourages the output of the encoder to stay close to the chosen codebook vector to prevent it from fluctuating too frequently from one code vector to another).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate minimizing a difference between a discrete latent sequence and a latent sequence output by the encoder module as taught by Razavi into the conditional VAE framework as taught by Wang. Doing so ensures the encoder outputs are well-aligned with discrete latent variables, therefore improving latent stability and reducing fluctuations in training (Razavi Pg. 2, Section 2.1 commitment loss… to prevent it from fluctuating too frequently from one code vector to another).
Regarding claim 5, Wang in view of Razavi discloses the motion generation system of claim 4, and further discloses wherein the instructions, when executed by the one or more processors, cause the one or more processors to (Pg. 7, Section 4.4; We train our model with a batch size of 32 on a V100 GPU) train the encoder module and the decoder module based on minimizing a difference between an output of the decoder module and a predetermined output (Wang Pg. 6, Section 4.3; Reconstruction loss…Eqn. (2)…distance between the predicted body parameters and ground truth (i.e., global translation t, global rotation r, and joint rotations p)).
Wang and Razavi are combined for the reason set forth above with respect to claim 4.
Claim(s) 7 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang et al. “HUMANISE: Language-conditioned human motion generation in 3d scenes”. In NeurIPS, 2022, hereinafter referred to as “Wang”, in view of Petrovich et al. “Action-conditioned 3d human motion synthesis with transformer vae”. In ICCV, 2021. (Year: 2021), hereinafter referred to as “Petrovich”.
Regarding claim 7, Wang discloses the motion generation system of claim 1, but does not disclose wherein the encoder module includes a Transformer architecture.
In the same art of conditioned 3D human motion with VAEs, Petrovich discloses wherein the encoder module includes a Transformer architecture (Pg. 1, Section 1; Transformer-based encoder-decoder architecture. Fig. 2; Transformer Encoder).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the Transformer encoder of Petrovich into Wang’s GRU encoder. Transformer-based architecture is a well-known architecture in sequence generation tasks such as human motion prediction. Incorporating a transformer-based encoder yields the predictable enhancements of better modeling of interaction with scenes and reduced motion regression (Pg. 1, Section 1; …this paves the way for better modeling of interaction with the environment…allows the generation of variable length sequences without the problems of motions regressing), leading to more accurate and realistic human motion renderings.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JENNY NGAN TRAN whose telephone number is (571)272-6888. The examiner can normally be reached Mon-Thurs 8am-5pm.
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, Alicia Harrington can be reached at (571) 272-2330. 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.
/JENNY N TRAN/Examiner, Art Unit 2615
/ALICIA M HARRINGTON/Supervisory Patent Examiner, Art Unit 2615