DETAILED ACTION
This Office action is in response to communications filed on 08/12/2025 for Application No. 17/855,729, in which claims 1-20 are presented for examination. The amendments filed on 08/12/2025 have been entered, where claims 1-3, 5-7, 8-10, 12-17, and 19-20 have been amended.
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 .
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(d):
(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
Claims 3, 10, and 17 are rejected under 35 U.S.C. 112(d) as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends.
Regarding Claim 3, the additional elements of the dependent claim are all previously recited in a claim upon which it depends (compare Claim 3, Ln. 1-4, “wherein the pairwise logistic loss function is modified such that loss terms associated with pairs of actors both ranked below a predetermined threshold in the training ranking are reduced or omitted, thereby focusing training on actors above the threshold” with Claim 1, Ln. 23-26, “wherein the pairwise logistic loss function is further modified such that, for a predetermined resource budget K, loss terms associated with pairs of actors both having a training rank below K are reduced or omitted, thereby focusing training on actors whose training rank is within the resource budget K”). An element-by-element comparison is provided below.
Claim 3
Claim 1
Examiner’s Comments
wherein the pairwise logistic loss function is modified such that
wherein the pairwise logistic loss function is further modified such that
loss terms associated with pairs of actors both ranked below a predetermined threshold in the training ranking are reduced or omitted
for a predetermined resource budget K, loss terms associated with pairs of actors both having a training rank below K are reduced or omitted
The predetermined resource budget K is a predetermined threshold, recited with additional detail.
thereby focusing training on actors above the threshold
thereby focusing training on actors whose training rank is within the resource budget K
Within and above are equivalent when recited in the context of reducing and or omitting actors below.
The focusing of claim 1 recites additionally details, but covers all elements of the focusing of claim 3.
Applicant may cancel the claim(s), amend the claim(s) to place the claim(s) in proper dependent form, rewrite the claim(s) in independent form, or present a sufficient showing that the dependent claim(s) complies with the statutory requirements.
Regarding Claim 10, the claim fails to further limit the subject matter of the claim upon which it depends for substantially the same reason as Claim 3. Therefore, it rejected under the same grounds. Applicant my respond in the same manner as described in regards to Claim 3.
Regarding Claim 17, the claim fails to further limit the subject matter of the claim upon which it depends for substantially the same reason as Claim 3. Therefore, it rejected under the same grounds. Applicant my respond in the same manner as described in regards to Claim 3.
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.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Refaat et al. (hereinafter Refaat) (“Agent Prioritization for Autonomous Navigation”) in view of Lee et al. (hereinafter Lee) (“DESIRE: Distant Future Prediction in Dynamic Scenes with Interacting Agents”) and Wang et al. (hereinafter Wang) (“The LambdaLoss Framework for Ranking Metric Optimization”).
Regarding Claim 1, Refaat teaches a method (Pg. 2062, Col. 2, Para. 2, “We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features”; Pg. 2060, Abstract, “We propose a system to rank agents around an autonomous vehicle (AV) in real time”) comprising:
identifying a set of actors in an environment of a vehicle (Pg. 2063, Col. 1-2, Para. 3-1, “The input to the deep net is a set of 10 top-down images as follows: . . . 6) Agents Image: the bounding boxes of all the agents located in 80 × 80 square meters around the AV at the current time”, where data must be identified to be used as “input”; see generally Pg. 2060, Col. 1, Para. 2, “An autonomous vehicle (AV) uses a variety of sensors such as radars, lidars and cameras to detect agents (e.g. other vehicles, cyclists and pedestrians) in the scene”) and
associated actor features for each actor (Pg. 2063, Col. 1-2, Para. 3-1, “The input to the deep net is a set of 10 top-down images as follows: . . . 4) Three images for other agents’ headings, velocities and accelerations, respectively, at any point in time. 5) Other Agents Time Image: the relative times of the agents in the pixels corresponding to where the agents were located, or are predicted to be located, at these relative times”, where “headings, velocities, accelerations, . . . and “locat[ions]” are within the broadest reasonable interpretation of features, which is for each agent, see Pg. 2063, Col. 1, Para. 2, “To create the input images, we use the history of all agents and the AV”);
identifying one or more intent features describing planned movement of the vehicle (Pg. 2063, Col. 1, Para. 3, “The input to the deep net is a set of 10 top-down images as follows: . . . 2) Three images for headings, velocities, and accelerations, respectively, of the AV at any point in time. 3) Time Image: every relative time of the AV in the pixel corresponding to where the AV was located, or is predicted to be located, at this relative time”, where “velocities, and accelerations” and information on where the vehicle is “predicted to be located” are within the broadest reasonable interpretation of intent features describing planned movement);
processing, by a neural network comprising one or more multi-layer perceptron (MLP) and/or convolutional neural network (CNN) layers, the actor features and the one or more intent features to generate [an output] . . . (Pg 2062, Col. 2, Para. 2, “Ranking Model . . . We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features . . . We next describe every component of our hybrid model”, where the “hybrid model” includes a “CNN”; Pg. 2064, Col. 1, Fig. 6, where the “network” uses multiple “layers” to process an “input” for use in generating the “output”, and where the “input” is the “top-down images”, which as discussed above, includes the actor features and intent features, see Pg. 2062, Col. 2, Para. 3, “Our ranking CNN takes 200 x 200 pixels top-down images as input . . . The output of the network is an image with importance scores in the pixels where the agents are located as shown in Figure 2”; see also Pg. 2063, Col. 1, Fig. 2);
for each actor, computing a learnable attention weight using a function of the . . . [CNN] (Pg. 2062, Col. 2, Para. 3, “The output of the network is an image with importance scores in the pixels where the agents are located as shown in Figure 2”, where the “importance scores” are within the broadest reasonable interpretation of attention weights because they are used to weight the level of attention the “agents” should receive during “the decision making process”, see Pg. 2060, Abstract, “With limited computational resources and a large number of agents to process in real time, it becomes important to efficiently rank agents according to their impact on the decision making process”, and where the “importance scores” are for each actor because they are used to “rank” each “agent” and correspond with the “locat[ion]” of the “agent”);
applying an attention model (Pg 2062, Col. 2, Para. 2, “Ranking Model . . . We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features . . . We next describe every component of our hybrid model”, where the “hybrid model” includes a “CNN”, which is an attention model, see Pg. 2060, Abstract, “With limited computational resources and a large number of agents to process in real time, it becomes important to efficiently rank agents according to their impact on the decision making process”; Pg. 2061, Col. 1, Para. 5, “We design a model based on gradient boosted decision trees (GBDT), which . . . optimizes a pairwise loss to predict an importance score for each agent”, where a model to predict an “importance score” for use in allocating “computational resources” based on “impact” is within the broadest reasonable interpretation of an attention model)
that receives the . . . [CNN output, including] the attention weights (Pg. 2063, Col. 1, Para. 3, “The input to the deep net is a set of 10 top-down images as follows . . . ”, where, as discussed above, the “deep net” is a component of the attention model and, as also discussed above, the “input to the deep net” includes actor and intent features, which is processed and input to the “GBDT”, see Pg. 2062, Col. 2, Para. 2, “We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features” and Pg. 2063, Col. 2, Para. 5, “The CNN score can be provided among the GBDT features”, where, as discussed above, the “CNN score” includes the attention weights)
to output a predicted ranking of the set of actors based on . . . the attention model (Pg. 2061, Col. 1, Para. 2, “We propose a system to rank agents around an AV in real time according to their importance to the decision making of the AV”; see also Pg. 2065, Table 2, where the “Pairwise CNN” and “Pairwise GBDT” in combination; see generally Pg. Col. 1, Para. 2, “A GBDT could be used for ranking by predicting a score for each agent. The final ranking is determined by these scores”); and
training . . . the neural network and the attention model using a pairwise logistic loss function determined by a relative ordering of the predicted ranking compared to a training ranking of the set of detected actors, wherein the training comprises . . . [updating] the neural network and attention model based on the pairwise ordering loss (Pg. 2063, Col. 2, Para. 4, “Specific losses for ranking typically yield superior results over classification losses [28]. As such, we use a pairwise logistic loss applied to pairs of agents. Let the labels of Agents i and j be li and lj , and their scores from the deep net be si and sj , respectively. The loss incurred by a pair is log (1 + esj−si ) if li > lj , and 0 otherwise”, where “pairwise logistic loss” is a function determined by the relative ordering compared with “labels”, which correspond with the training ranking for use in training the CNN, which requires updating the CNN; Pg. 2064, Col. 2, Para. 3, “To train the GBDT for ranking, we used the standard logistic loss over the difference of predictions for pairs of agents [28]”, where “train[ing] the GBDT”, which requires updating the GBDT, uses a “logistic loss” function determined by “pairs of agents”, see generally Pg. 2062, Col. 1, Para. 2, “A common approach to train the GBDT for ranking is to transform a list of objects to be ranked into a set of object pairs, and optimize a pairwise loss that penalizes the misorder of each pair with respect to the ground truth ranking; see [28]”, where the method of “[28]”, which is used by Refaat therein, is a relative ordering - “pairwise loss”, compared to a training ranking - “ground truth ranking”),
wherein the pairwise logistic loss function . . . [applied to] pairs of actors . . . on actors . . . (Pg. 2063, Col. 2, Para. 4, “Specific losses for ranking typically yield superior results over classification losses [28]. As such, we use a pairwise logistic loss applied to pairs of agents. Let the labels of Agents i and j be li and lj , and their scores from the deep net be si and sj , respectively. The loss incurred by a pair is log (1 + esj−si ) if li > lj , and 0 otherwise”).
Refaat does not explicitly disclose . . . for each actor, a respective actor embedding and to generate a scene embedding based on the one or more intent features and a joint actor embedding formed from a combination of the actor embeddings . . . respective actor embedding and the scene embedding . . . actor embeddings, the scene embedding . . . a set of parameters . . . the parameters of . . . performing back-propagation to update the weights of . . . is further modified such that, for a predetermined resource budget K, loss terms associated with . . . both having a training rank below K are reduced or omitted, thereby focusing training . . . whose training rank is within the resource budget K.
However, Lee teaches . . . [a model to generate] for each actor, a respective actor embedding (Pg. 336, Col. 1, Abstract, “We introduce a Deep Stochastic IOC1 RNN Encoder-decoder framework, DESIRE, for the task of future predictions of multiple interacting agents in dynamic scenes”, where the “RNN” generates an actor embedding, “HXi and HYi”, see Pg. 339, Col. 1, Para. 2, “the past and future trajectories of an agent i, Xi and Yi respectively, are encoded through two RNN encoders with separate set of parameters (i.e., RNN Encoder1 and RNN Encoder2 in Fig. 2). The resulting two encodings, HXi and HYi , are concatenated”; Pg. 341, Col. 1, Para. 2, “The embedding vector HXi (the output of the RNN Encoder1 in Fig. 2)”, where “embedding” and “encodings” are used interchangeably)
and to generate a scene embedding based on the one or more intent features (Pg. 337, Col. 1, Para. 1, “the scene context encoded by a convolutional neural network (CNN)”, where “the scene context” is based on one or more intent features, see Pg. 366, Fig. 1, where the “driving scenario” which is the intended direction of the point-of-view car; see also Pg. 337, Col. 1, Para. 1, “scene context derived from image-based features or other sensory data if available”, where “scene context” is “derived from . . . sensory data”, which in turn is based on the intent features of direction)
and a joint actor embedding formed from a combination of the actor embeddings (Pg. 337, Col. 1, Para. 1, “Scene Context Fusion: Sec. 3.3 presents the Scene Context Fusion (SCF) layer that aggregates interactions between agents and the scene context encoded by a convolutional neural network (CNN). The fused embedding is channeled to the aforementioned RNN scoring module and allows to produce the rewards based on the contextual information”, where the “Scene Context Fusion” includes the combination of “agents” as a “fused embedding”, which is a joint actor embedding)
. . . [use of the] respective actor embedding and the scene embedding [to generate a value] (Pg. 340, Col. 1-2, Para. 3-1, “Let the score s of individual prediction hypothesis . . . be defined as [equation 1] . . . the reward function ψ incorporates both scene context I as well as the interaction between agents”)
[and providing a model with the] . . . actor embeddings, [and] the scene embedding . . . (Pg. 336, Col. 1, Abstract, “We introduce a Deep Stochastic IOC1 RNN Encoder-decoder framework”, where the “decoder” receives the embeddings, see Pg. 339, Fig. 2; see also Pg. 341, Col. 1, Para. 1, “The embedding vector HXi (the output of the RNN Encoder1 in Fig. 2) is shared as the initial hidden state of the RNN, in order to provide the individual past motion context . . . Combined with CNN features, the SCF module provides the RNN decoder with both static and dynamic scene information. It learns consistency between semantics of agents and scenes for reliable prediction”)
. . . [generate an output based on] a set of parameters of [multiple models] (Pg. 336, Col. 1, Abstract, “We introduce a Deep Stochastic IOC1 RNN Encoder-decoder framework, DESIRE, for the task of future predictions of multiple interacting agents in dynamic scenes. DESIRE effectively predicts future locations of objects in multiple scenes”, where the “predict[ions]” are based on parameters, see Pg. 339, Col. 1, Para. 1, “Here, θ, φ, ν denote the parameters of corresponding networks”; Pg. 340, Col. 2, Para. 2, “Learning to refine: Alongside the scores, our model also estimates a regression vector . . . Represented as parameters of a neural network, the regression function η accumulates both scene contexts and all other agents dynamics from the past to entire future frames, and estimates the best displacement vector”)
. . . [training] the parameters of [the model] . . . [by] performing back-propagation to update the weights of [the model] . . . (Pg. 339, Col. 2, Para. 1, “Since back-propagation is not possible through random sampling, we adopt the standard reparameterization trick [22] to make it differentiable”, where “back-propagation” requires adjustment of weight values to train the model and where the entire network is “trainable” from “end-to-end”, see Pg. 336, Col. 1, Abstract, “DESIRE achieves these in a single end-to-end trainable neural network model”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the identification of actors, associated actor features, and intent features in an environment; the use of a multi-layer convolutional neural network to produce an output based on the identified features, where the output includes an attention weight for each actor; the use of an attention model to predict a ranking of actors based on the convolutional neural network output; and the training of the convolutional neural network and attention model using a pairwise logistic loss function determined by the relative ordering of the pairs in the predicted ranking of Refaat with the use of a model to generate actor, scene, and joint actor embeddings, where the scene embedding is based on intent features and the joint actor embedding comprises the actor embeddings; the use of the embeddings to generate a value; the provision of the embeddings to a model; the generation of an output based on a set of parameters of multiple models; and the training of parameters using backpropagation to update model weights of Lee in order to coherently integrate dynamic data from actors and scenes into model predictions, where training of model parameters using backpropagation to adjust weights corresponding to model outputs is a well-established and understood training approach (Lee, Pg. 336, Col. 1, Abstract, “DESIRE effectively predicts future locations of objects in multiple scenes by 1) accounting for the multi-modal nature of the future prediction (i.e., given the same context, future may vary), 2) foreseeing the potential future outcomes and make a strategic prediction based on that, and 3) reasoning not only from the past motion history, but also from the scene context as well as the interactions among the agents”; Lee, Pg. 340, Col. 1, Para. 2, “Our model learns the agents dynamics as well as the scene context in a coherent framework”; Lee, Pg. 341, Col. 1, Para. 2, “Instead of using the max pooling operation with rectangular grids, we adopt log-polar grids with an average pooling. Combined with CNN features, the SCF module provides the RNN decoder with both static and dynamic scene information”), which offers a clear improvement over baselines (Lee, Pg. 343, Col. 2, Para. 2, “Our empirical evaluations on driving and surveillance scenarios demonstrate clear improvement over other baselines”).
Additionally, Wang teaches . . . [a pairwise logistic loss function] (Pg. 1315, Col. 1, Para. 3, “A commonly used pairwise loss function is the logistic loss [equation 3] . . . Note that we use log2 instead of log”)
is further modified such that, for a predetermined resource budget K, loss terms associated with . . . [pairs of entities] both having a training rank below K are reduced or omitted, thereby focusing training on . . . [entities] whose training rank is within the resource budget K (Pg. 1315, Col. 2, Para. 3, “LambdaRank uses the logistic loss in Eq 3 and adapts it by reweighing each document pair by ∆NDCG in each iteration [equation 6]”, where “LambdaLoss” modifies the “logistic loss” so that “document pair[s]” where both entities rank below the top k values, meaning there numeric ranking is larger than resource budget k, “i ≤ k or j ≥ k” not satisfied when the value is greater than predetermined “NDCG@5”, the loss will be “truncat[ed]” to omit the pair’s loss, thereby focusing on pairs with a value in the resource budget, see Pg. 1315, Col. 2, Para. 5, “we present the LambdaLoss framework and provide an underlying loss for LambdaRank by formulating it as a special configuration in the LambdaLoss framework”; Pg. 1319, Col. 2, Para. 3, “we use NDCG@5 as our primary metric. . . For our metric-driven loss functions, we adapt them for NDCG@k by truncating the loss. For example, we have NDCG-Loss2 variant for NDCG@k as [equation 18] . . . In our experiments, when NDCG@k is used as the evaluation metric, we use the corresponding truncated LambdaLoss”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the training of a neural network and attention model using a pairwise logistic loss function determined by the relative ordering of the predicted ranking compared with a training ranking of Refaat in view of Lee with the use of a modified pairwise logistic loss function to omit pairs where both entities are below a predefined resource budget, thereby focusing on entities within a resource budget of Wang in order to focus training loss on outputs more relevant to end users (Wang, Pg. 1313, Col. 1-2, Para. 3-1, “Since users, when presented with a ranked list of documents, are more likely to scan documents downwards starting at the top, most ranking metrics are rank-dependent and reward relevance of top-ranked documents more”) and to establish consistent practices across training and evaluation (compare Wang, Pg. 1319, Col. 2, Para. 3, “In our experiments, when NDCG@k is used as the evaluation metric, we use the corresponding truncated LambdaLoss” with Refaat, Pg. 2065, Col. 1-2, Para. 2-1, “The training set consists of planner reactions for agents from ∼ 54 million planning iterations . . . we would like a metric that favors the appearance of highly relevant agents earlier in the sorted list. We use the NDCG [16] which penalizes the appearance of a highly relevant agent in lower positions as the graded relevance value is reduced logarithmically proportional to the position in the sorted list; see [34]. To compute the NDCG, we first assign a numerical score r(k) to every k th position of the sorted list . . . 2 for most relevant agents, 1 for relevant agents and 0 for less relevant agents. The discounted cumulative gain (DCG) for the first K agents in a sorted list is computed . . . This way DCG@K penalizes putting highly relevant agents at lower positions in a sorted list”).
Regarding Claim 2, Refaat in view of Lee and Wang teach the method of claim 1, wherein the pairwise logistic loss function comprises, for each pair of actors, a loss term that penalizes incorrect ordering of the predicted ranking relative to the training ranking (Refaat, Pg. 2063, Col. 2, Para. 4, “we use a pairwise logistic loss applied to pairs of agents. Let the labels of Agents i and j be li and lj , and their scores from the deep net be si and sj , respectively. The loss incurred by a pair is log (1 + esj−si ) if li > lj , and 0 otherwise”, where the first and second actor are “i” and “j” respectively and the loss is relative to the “labels”, which correspond with the training ranking and the term that penalizes incorrect ordering of the predicted ranking is “log (1 + esj−si ) if li > lj”).
Regarding Claim 3, Refaat in view of Lee and Wang teach the method of claim 1, wherein the pairwise logistic loss function is modified such that loss terms associated with pairs of actors both ranked below a predetermined threshold in the training ranking are reduced or omitted, thereby focusing training on actors above the threshold (Refaat, Pg. 2063, Col. 2, Para. 4, “we use a pairwise logistic loss applied to pairs of agents. Let the labels of Agents i and j be li and lj , and their scores from the deep net be si and sj , respectively. The loss incurred by a pair is log (1 + esj−si ) if li > lj , and 0 otherwise”, where, in view of Wang, the loss function is modified omit pairs where both entities are below the threshold, see Wang, Pg. 1315, Col. 2, Para. 3, “LambdaRank uses the logistic loss in Eq 3 and adapts it by reweighing each document pair by ∆NDCG in each iteration [equation 6]”, where “LambdaLoss” modifies the “logistic loss” so that “document pair[s]” where both entities rank below the top k values, meaning there numeric ranking is larger than resource budget k, “i ≤ k or j ≥ k” not satisfied when the value is greater than predetermined “NDCG@5”, the loss will be “truncat[ed]” to omit the pair’s loss, thereby focusing on pairs with a value in the resource budget, see Wang, Pg. 1315, Col. 2, Para. 5, “we present the LambdaLoss framework and provide an underlying loss for LambdaRank by formulating it as a special configuration in the LambdaLoss framework”; Wang, Pg. 1319, Col. 2, Para. 3, “we use NDCG@5 as our primary metric. . . For our metric-driven loss functions, we adapt them for NDCG@k by truncating the loss. For example, we have NDCG-Loss2 variant for NDCG@k as [equation 18] . . . In our experiments, when NDCG@k is used as the evaluation metric, we use the corresponding truncated LambdaLoss”).
The reasons of obviousness have been noted in the rejection of claim 1 above and remain applicable here.
Regarding Claim 4, Refaat in view of Lee and Wang teach the method of claim 1, wherein the actor features and one or more intent features include features at a plurality of times (Refaat, Pg. 2063, Fig. 4, “We provide a separate time channel for each of the AV and the Agents”; Refaat, Pg. 2063, Fig. 3, “The correspondence between the time channel and other input channels”, where the “time channel” corresponds with input values, which as discussed above includes actor features and intent features, at a plurality of times; see also Refaat, Pg. 2063, Col. 1, Para. 1, “For example, Figure 3 shows how the time channel can recover information about acceleration over time when combined with the
acceleration channel . . . Time in the past is represented with negative numbers, whereas future time is represented with positive numbers”; see generally Refaat, Pg. 2061, Col. 1, Para. 4, “Data Representation: We represent perceptual outputs through time (e.g. agent locations, speed and acceleration profiles) compactly to provide input to a deep convolutional neural network . . . The main idea is to represent the time information as a separate channel which, when combined with other input channels, provides information about the scene as it changes through time”, which includes actor features and where intent features must also be included at a plurality of times in order to “augment” inputs based on the “time at which the AV's trajectory will be the closest to the agent”, see Refaat, Pg. 2064, Col. 1, Para. 1).
Regarding Claim 5, Refaat in view of Lee and Wang teach the method of claim 1, wherein the attention model is configured to (Refaat, Pg 2062, Col. 2, Para. 2, “Ranking Model . . . We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features”):
generate, by the neural network, a set of actor embeddings corresponding to the set of detected actors based on respective actor features for each actor (Refaat, Pg. 2063, Col. 1-2, Para. 3-1, “The input to the deep net is a set of 10 top-down images as follows: . . . 4) Three images for other agents’ headings, velocities and accelerations, respectively, at any point in time. 5) Other Agents Time Image: the relative times of the agents in the pixels corresponding to where the agents were located, or are predicted to be located, at these relative times”, where “headings, velocities, accelerations, . . . and “locat[ions]” are within the broadest reasonable interpretation of features, which is for each agent, see Refaat, Pg. 2063, Col. 1, Para. 2, “To create the input images, we use the history of all agents and the AV”, and where, in view of Lee, the features are used by the neural network to generate the actor embeddings, see Lee, Pg. 339, Col. 1, Para. 2, “the past and future trajectories of an agent i, Xi and Yi respectively, are encoded through two RNN encoders with separate set of parameters (i.e., RNN Encoder1 and RNN Encoder2 in Fig. 2). The resulting two encodings, HXi and HYi , are concatenated”, where “past and future trajectories” are features; Lee, Pg. 336, Col. 1, Abstract, “We introduce a Deep Stochastic IOC1 RNN Encoder-decoder framework, DESIRE, for the task of future predictions of multiple interacting agents in dynamic scenes”, where the predictions, and therefore the embeddings, are generated for the “agents” in the scene, which is the set of actors);
generate a scene embedding based on the one or more intent features and a joint actor embedding formed from a combination of the actor embeddings (Refaat, Pg. 2063, Col. 1, Para. 3, “The input to the deep net is a set of 10 top-down images as follows: . . . 2) Three images for headings, velocities, and accelerations, respectively, of the AV at any point in time. 3) Time Image: every relative time of the AV in the pixel corresponding to where the AV was located, or is predicted to be located, at this relative time”, where “velocities, and accelerations” and information on where the vehicle is “predicted to be located” are within the broadest reasonable interpretation of intent features describing planned movement, which in view of Lee is used to generate the scene embedding; Lee, Pg. 337, Col. 1, Para. 1, “Scene Context Fusion: Sec. 3.3 presents the Scene Context Fusion (SCF) layer that aggregates interactions between agents and the scene context encoded by a convolutional neural network (CNN). The fused embedding is channeled to the aforementioned RNN scoring module and allows to produce the rewards based on the contextual information”, where the “Scene Context Fusion” includes the combination of “agents” as a “fused embedding”, which is a joint actor embedding, and is also a “scene” embedding because it is based on “the scene context”, which as discussed above, includes intent features, see Lee, Pg. 366, Fig. 1, where the “driving scenario” which is the intended direction of the point-of-view car; see also Lee, Pg. 337, Col. 1, Para. 1, “scene context derived from image-based features or other sensory data if available”, where “scene context” is “derived from . . . sensory data”, which in turn is based on the intent features of direction);
and compute, for each actor, the learnable attention weight using a function of the respective actor embedding and the scene embedding, and determine the respective rank of each actor based on the attention weights (Refaat, Pg. 2062, Col. 2, Para. 3, “The output of the network is an image with importance scores in the pixels where the agents are located as shown in Figure 2”, where the “importance scores” are within the broadest reasonable interpretation of attention weights because they are used to weight the level of attention the “agents” should receive during “the decision making process”, see Refaat, Pg. 2060, Abstract, “With limited computational resources and a large number of agents to process in real time, it becomes important to efficiently rank agents according to their impact on the decision making process”, and where the “importance scores” are for each actor because they are used to “rank” each “agent” and correspond with the “locat[ion]” of the “agent”, which, in view of Lee, the output of the CNN used to generate the attention weights are the actor and scene embedding, see Lee, Pg. 337, Col. 1, Para. 1, “the fused embedding is channeled to the aforementioned RNN scoring module” and Lee, Pg. 340, Col. 1-2, Para. 3-1, “Let the score s of individual prediction hypothesis . . . be defined as [equation 1] . . . the reward function ψ incorporates both scene context I as well as the interaction between agents”).
The reasons of obviousness have been noted in the rejection of claim 1 above and remain applicable here.
Regarding Claim 6, Refaat in view of Lee and Wang teach the method of claim 5, wherein the scene embedding is based on a joint actor embedding determined by combining the set of actor embeddings (Lee, Pg. 337, Col. 1, Para. 1, “Scene Context Fusion: Sec. 3.3 presents the Scene Context Fusion (SCF) layer that aggregates interactions between agents and the scene context encoded by a convolutional neural network (CNN). The fused embedding is channeled to the aforementioned RNN scoring module and allows to produce the rewards based on the contextual information”, where the “Scene Context Fusion” includes the combination of “agents” as a “fused embedding”, which is a joint actor embedding, and is also a “scene” embedding because it is based on “the scene context”, which as discussed above, includes intent features, see Lee, Pg. 366, Fig. 1, where the “driving scenario” which is the intended direction of the point-of-view car)
using a permutation-invariant operation selected from the group consisting of summation, averaging, or max pooling (Lee, Pg. 341, Col. 1, Para. 3, “we pool the hidden representation of all the other agents’ samples that are within the spatial cell, ∀j ≠ i, ∀k, yˆ(k) j,t ∈ g. Instead of using the max pooling operation with rectangular grids, we adopt log-polar grids with an average pooling”, where “average pooling” is permutation-invariant because it is order independent).
The reasons of obviousness have been noted in the rejection of claim 1 above and remain applicable here.
Regarding Claim 7, Refaat in view of Lee and Wang teach the method of claim 5, wherein the parameters of the attention model comprise weights of the neural network layers used to generate the actor embeddings, the joint actor embedding, and the scene embedding (Refaat, Pg 2062, Col. 2, Para. 2, “Ranking Model . . . We propose a hybrid approach which trains a GBDT on a combination of engineered and learned CNN features . . . We next describe every component of our hybrid model”, where the “hybrid model” includes a “CNN”; Refaat, Pg. 2064, Col. 1, Fig. 6, where the “network” uses multiple “layers” to process an “input” for use in generating the “output”, which, in view of Lee includes is the generated embeddings, see Lee Pg. 339, Col. 1, Para. 2, “the past and future trajectories of an agent i, Xi and Yi respectively, are encoded through two RNN encoders with separate set of parameters (i.e., RNN Encoder1 and RNN Encoder2 in Fig. 2). The resulting two encodings, HXi and HYi , are concatenated”, Lee, Pg. 337, Col. 1, Para. 1, “the scene context encoded by a convolutional neural network (CNN)”, and Lee, Pg. 337, Col. 1, Para. 1, “Scene Context Fusion: Sec. 3.3 presents the Scene Context Fusion (SCF) layer that aggregates interactions between agents and the scene context encoded by a convolutional neural network (CNN). The fused embedding is channeled to the aforementioned RNN scoring module and allows to produce the rewards based on the contextual information”, which are generated using parameter weights, see Lee, Pg. 336, Col. 1, Abstract, “We introduce a Deep Stochastic IOC1 RNN Encoder-decoder framework, DESIRE, for the task of future predictions of multiple interacting agents in dynamic scenes. DESIRE effectively predicts future locations of objects in multiple scenes”, where the “predict[ions]” are based on parameters, see Lee, Pg. 339, Col. 1, Para. 1, “Here, θ, φ, ν denote the parameters of corresponding networks”; Lee, Pg. 340, Col. 2, Para. 2, “Learning to refine: Alongside the scores, our model also estimates a regression vector . . . Represented as parameters of a neural network, the regression function η accumulates both scene contexts and all other agents dynamics from the past to entire future frames, and estimates the best displacement vector”),
and wherein the weights are updated during training by back-propagation (Lee, Pg. 339, Col. 2, Para. 1, “Since back-propagation is not possible through random sampling, we adopt the standard reparameterization trick [22] to make it differentiable”, where “back-propagation” requires adjustment of weight values to train the model and where the entire network is “trainable” from “end-to-end”, see Lee, Pg. 336, Col. 1, Abstract, “DESIRE achieves these in a single end-to-end trainable neural network model”)
based on the pairwise logistic loss (Refaat, Pg. 2063, Col. 2, Para. 4, “Specific losses for ranking typically yield superior results over classification losses [28]. As such, we use a pairwise logistic loss applied to pairs of agents. Let the labels of Agents i and j be li and lj , and their scores from the deep net be si and sj , respectively. The loss incurred by a pair is log (1 + esj−si ) if li > lj , and 0 otherwise”, where “pairwise logistic loss” is a function determined by the relative ordering compared with “labels”, which correspond with the training ranking for use in training the CNN, which requires updating the CNN; Refaat, Pg. 2064, Col. 2, Para. 3, “To train the GBDT for ranking, we used the standard logistic loss over the difference of predictions for pairs of agents [28]”, where “train[ing] the GBDT”, which requires updating the GBDT, uses a “logistic loss” function determined by “pairs of agents”, see generally Refaat, Pg. 2062, Col. 1, Para. 2, “A common approach to train the GBDT for ranking is to transform a list of objects to be ranked into a set of object pairs, and optimize a pairwise loss that penalizes the misorder of each pair with respect to the ground truth ranking; see [28]”, where the method of “[28]”, which is used by Refaat therein, is a relative ordering - “pairwise loss”, compared to a training ranking - “ground truth ranking”).
The reasons of obviousness have been noted in the rejection of claim 1 above and remain applicable here.
Regarding Claim 8 Refaat teaches a system, comprising: a processor; and a non-transitory computer-readable storage medium containing instructions for execution by the processor for: . . . (Refaat, Pg. 2066, Col. 2, Footnote 5, “In our experiments, rendering is done on an Intel Xeon W-2135 processor, and Nvidia Tesla V100 is used for inference”, where the use of a ”processor” to execute the method requires the processor be part of a system with a non-transitory computer-readable medium containing instructions, because a processor cannot execute a method without being part of a system with instructions that must be stored in a non-transitory computer-readable medium).
The remaining limitations are substantially the same as limitations of Claim 1, therefore it is rejected under the same rational.
Regarding Claim 9, the additional elements of the dependent claim are substantially the same as limitations of Claim 2, therefore it is rejected under the same rationale.
Regarding Claim 10, the additional elements of the dependent claim are substantially the same as limitations of Claim 3, therefore it is rejected under the same rationale.
Regarding Claim 11, the additional elements of the dependent claim are substantially the same as limitations of Claim 4, therefore it is rejected under the same rationale.
Regarding Claim 12, the additional elements of the dependent claim are substantially the same as the limitations of Claim 5, therefore it is rejected under the same rationale.
Regarding Claim 13, the additional elements of the dependent claim are substantially the same as the limitations of Claim 6, therefore it is rejected under the same rationale.
Regarding Claim 14, the additional elements of the dependent claim are substantially the same as the limitations of Claim 7, therefore it is rejected under the same rationale.
Regarding Claim 15 Refaat teaches a non-transitory computer-readable medium containing instructions executable by a processor for . . . (Refaat, Pg. 2066, Col. 2, Footnote 5, “In our experiments, rendering is done on an Intel Xeon W-2135 processor, and Nvidia Tesla V100 is used for inference”, where the use of a ”processor” to execute the method requires a non-transitory computer-readable medium containing instructions, because a processor cannot execute a method without instructions that must be stored in a non-transitory computer-readable medium).
The remaining limitations are substantially the same as limitations of Claim 1, therefore it is rejected under the same rational.
Regarding Claim 16, the additional elements of the dependent claim are substantially the same as limitations of Claim 2, therefore it is rejected under the same rationale.
Regarding Claim 17, the additional elements of the dependent claim are substantially the same as limitations of Claim 3, therefore it is rejected under the same rationale.
Regarding Claim 18, the additional elements of the dependent claim are substantially the same as limitations of Claim 4, therefore it is rejected under the same rationale.
Regarding Claim 19, the additional elements of the dependent claim are substantially the same as the limitations of Claim 5, therefore it is rejected under the same rationale.
Regarding Claim 20, the additional elements of the dependent claim are substantially the same as the limitations of Claim 6, therefore it is rejected under the same rationale.
Response to Arguments
Applicant's arguments filed on 08/12/2025 have been fully considered, but all are rendered moot by Applicant’s amendments. Additional details are provided below.
I. Applicant argues the rejections of claims 1-20, as being ineligible under U.S.C. § 101, are improper because the claims, as amended, are directed to patent eligible subject matter (Remarks, 08/12/2025, Pg. 13-14, Section “Rejections Under 35 U.S.C. § 101”).
In response to Applicant’s amendments, the previously communicated rejections under 35 U.S.C. § 101, have been withdrawn. As a result, the argument is rendered moot.
II. Applicant argues the rejections of claims 1-20, as being rendered obvious by prior art under 35 U.S.C. § 103, because the claims, as amended, are not rendered obvious by the prior art of record (Remarks, 08/12/2025, Pg. 14-16, Section “Rejections Under 35 U.S.C. § 103”).
In response to Applicant’s amendments, the previously communicated rejections under 35 U.S.C. § 103, have been withdrawn. However, Applicants arguments are not persuasive in light of the new rejection, under 35 U.S.C. § 103, discussed in detail above. The new grounds of rejection add new prior art to the art to the record in order to teach the new elements in the amended claims. the new grounds of rejection do not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. As a result, the argument is rendered moot.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 MATTHEW BRYCE GOLAN whose telephone number is (571)272-5159. The examiner can normally be reached Monday through Friday, 8:00 AM to 5:00 PM 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, Alexey Shmatov can be reached at (571) 270-3428. 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.
/MATTHEW BRYCE GOLAN/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights