DETAILED ACTION
This communication is in response to the amendment of 2/4/26 in which claims 1, 6, 10, 14, 18 were amended, and claim 20 was newly presented for examination. Claims 1-20 are pending.
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 .
Allowable Subject Matter
Claim 20 is objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Response to Arguments
Applicant’s arguments were considered. However, Wang is remapped to teach the amended limitations. Examiner respectfully directs Applicant to the detailed rejection below for explanation.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 6-13 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Wang, Jingkang, et al. "Advsim: Generating safety-critical scenarios for self-driving vehicles." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) (“Wang”).
Regarding claim 6, Wang discloses [a] method comprising:
receiving log data associated with a vehicle traversing a real-world environment, the log data associated with a first value of a first parameter associated with the environment; (see Wang pg. 9910, 1st column (“In this paper, we leverage real world traffic scenarios available in standard self-driving datasets…”); see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature.”); see Wang pg. 9914, Section 4.1 (“Dataset We evaluate our approach on a self-driving dataset, UrbanScenarios, which has 5,000 driving logs of 25 seconds each. Our dataset is collected across multiple cities in North America, and contains different types of map layouts and varying traffic densities.”))
inputting the log data into a search algorithm; (see Wang pg. 9913, 2nd column (“AdvSim is a framework that can use any black-box search algorithm to identify autonomy system failures. The search algorithm attempts to find the safety critical scenarios by maximizing the adversarial objective Ladv in Eq. 2. The search algorithm queries the autonomy system with a candidate perturbation τadv to obtain a query pair (τadv,Ladv) and maintains a history H of past uery pairs to generate the next candidate perturbation. We study a wide variety of black-box search algorithms including (1) Bayesian optimization [40, 35] (BO), (2) genetic algorithms [2] (GA), (3) random search [17] (RS) and (4) gradient estimation methods (NES [19] and Bandit-TD [20]).”))
receiving, from the search algorithm, a first modified value of the first parameter, (see Wang pg. 9910 2nd column (“There are three main components for generating safety-critical scenarios: a scenario parameterization space to optimize over, a search algorithm that identifies critical scenario parameters…”), pg. 9911 2nd column (“Our approach, AdvSim, works as follows: we first perturb the actors’ motion trajectories in an existing scenario…”), pg. 9912 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}T t=0 as a sequence of kinematic bicycle model states st = {xt,yt,θt,vt,κt,at} in the next T timesteps. Here (x,y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature. Candidate adversary trajectories can be generated by perturbing the change of curvature ˙κt and acceleration values at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”), pg. 9913, 1st column (“Since we aim for a general adversarial scenario generation framework, we consider the autonomy system as a black box, where we access the evaluation scores through limited queries. Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost. In this section, we introduce the adversarial objective we optimize to produce worst-case scenarios and detail the search algorithms applied.”)) wherein the search algorithm determines the first modified value of the first parameter based on a first simulation score, (see Wang pg. 9911 (“Our approach, AdvSim, works as follows: we first perturb the actors’ motion trajectories in an existing scenario, and generate the sequence of LiDAR point clouds that reflect the change in actor locations. With the adjusted sensor data, we run the autonomy stack [e.g., simulation] and get the planned SDV motion path. Finally, we evaluate the output path with a proposed adversarial objective and adjust the scenario perturbation to be more challenging.”), pg. 9913, 1st column (“Since we aim for a general adversarial scenario generation framework, we consider the autonomy system as a black box, where we access the evaluation scores [e.g., first simulation score] through limited queries.”), pg. 9914 1st column (“we update the sensor data accordingly (L.7) and evaluate the full autonomy system on generated scenarios to compute Ladv (L.8-9).”)
PNG
media_image1.png
510
512
media_image1.png
Greyscale
pg. 9913 2nd column (“AdvSim is a framework that can use any black-box search algorithm to identify autonomy system failures. The search algorithm attempts to find the safety critical scenarios by maximizing the adversarial objective Ladv in Eq. 2. The search algorithm queries the autonomy system with a candidate perturbation τadv to obtain a query pair (τadv,Ladv) and maintains a history H of past query pairs to generate the next candidate perturbation. We study a wide variety of black-box search algorithms including (1) Bayesian optimization [40, 35] (BO), (2) genetic algorithms [2] (GA), (3) random search [17] (RS) and (4) gradient estimation methods (NES [19] and Bandit-TD [20]).”)) the first simulation score associated with occurrence of a first event in a first run of a driving simulation, (see Wang pg. 9913, 2nd column (“Our use of multiple different costs allows us to identify different types of autonomy system failures, such as unnatural trajectories, collisions, and hard braking.”)) wherein the first run of the driving simulation is executed based at least in part on determining that the first modified value satisfies a simulation condition (Wang pg. 9912 2nd column (“To increase the perturbed trajectory’s plausibility, we ensure it does not collide with other actors or the original expert trajectory of the SDV [e.g., simulation condition]. In practice, we do this by first performing rejection sampling to create a set of physically feasible trajectories Tadv and then projecting the trajectory generated by δ on to the physically feasible set, measured by L2 distance. Our search space is low-dimensional and conducive to query-based black box optimization, while still allowing for fine-grained actor motion control.”)) the first run of the driving simulation including an autonomous vehicle and based at least in part on the first modified value of the first parameter (see Wang pg. 9913, 1st column (“Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost.”));
outputting first modified log data based at least in part on the first modified value of the first parameter; and (see Wang pg. 9910, 1st column (“As our perturbation modifies the actors’ trajectories, we need to adjust the sensor data to accurately reflect the actors’ new locations. We therefore adopt a high-fidelity LiDAR simulator [27] that modifies the sensor data accordingly taking into account occlusions.”))
training a machine learning model for controlling an autonomous vehicle based at least in part on the first modified log data (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors. We demonstrate the flexibility and scalability of our approach by generating over 4000 adversarial scenarios for a wide range of modern autonomy systems. Finally, we leverage AdvSim-generated safety-critical scenarios in training and further improve the safety of autonomy systems.”)).
Regarding claim 7, Wang discloses the invention of claim 6 as discussed above. Wang further discloses wherein the first parameter is associated with a first object in the environment, and wherein during the first run of the driving simulation a first simulated object is controlled to substantially correspond with the first object in the environment based on the log data and the first modified value of the first parameter throughout the driving simulation (see Wang pg. 9913, 1st column (“Once we have removed the selected actors from the LiDAR sweep, we update the LiDAR with the actors at their new locations. We first render the simulated LiDAR for the actors at their new locations using LiDAR-sim’s vehicle asset bank (Fig. 3e). Fig. 3f shows the real LiDAR point cloud with the added actors. However, when a LiDAR ray hits an object, the remaining path of the ray becomes occluded, creating a LiDAR shadow. Similar to the actor removal process, we create range images of the simulated and real LiDAR, and merge the LiDAR point clouds, thereby removing the LiDAR points of the now-occluded regions (Fig. 3g) and obtaining the final modified LiDAR sweep (Fig. 3h). The generated scenes are realistic and match the desired perturbation in actors’ motions (Fig. 3).”)).
Regarding claim 8, Wang discloses the invention of claim 6 as discussed above. Wang further discloses wherein the first parameter is associated with a first object in the environment, and wherein the first run of the driving simulation is executed in a first time period and a second time period, wherein during the first time period a first simulated object is controlled to substantially correspond with the first object in the environment based on the log data and the first modified value of the first parameter, (see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature.”)) and during the second time period the first simulated object is controlled using a planning component (see Wang pg. 9913, 2nd column (“Candidate adversary trajectories can be generated by perturbing the change of curvature κ˙ t and acceleration values at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”); see Wang pg. 9910, 2nd column (“Recently, interpretable neural motion planners [48, 49, 36] provide an alternative that inherits the advantages of traditional pipelines and end-to-end approaches, by maintaining modularity and interpretability while enabling end-to-end learning. This first began with joint perception and prediction [25, 10], which neural planners extended to include planning. Specifically, NMP [48] shared feature representations between multiple subtasks and predicted a cost volume to represent the quality of possible locations in planning. DSDNet [49] proposed an energy-based model to parameterize the joint distribution of the actors’ future trajectories. P3 [36] developed a semantic occupancy representation and generated consistent ego-vehicle plans.”); see Wang pg. 9911, 1st column (“We choose to represent the behavior of actors as kinematic bicycle-model trajectories, allowing for physical feasibility and fine-grained behavior control.”)).
Regarding claim 9, Wang discloses the invention of claim 6 as discussed above. Wang further discloses determining a search range defining permitted values of the first parameter, and providing the search range to the search algorithm as a constraint (see Wang pg. 9912, 2nd column (“To increase the perturbed trajectory’s plausibility, we ensure it does not collide with other actors or the original expert trajectory of the SDV. In practice, we do this by first performing rejection sampling to create a set of physically feasible trajectories Tadv and then projecting the trajectory generated by on to the physically feasible set, measured by L2 distance. Our search space is low-dimensional and conducive to query-based black box optimization, while still allowing for fine-grained actor motion control.”)).
Regarding claim 10, Wang discloses the invention of claim 9 as discussed above. Wang further discloses determining the search range based at least in part on map data associated with the environment, an environmental attribute associated with the environment, or values of the first parameter in additional log data associated with additional objects in the environment (see Wang pg. 9912, 2nd column (“To increase the perturbed trajectory’s plausibility, we ensure it does not collide with other actors or the original expert trajectory of the SDV. In practice, we do this by first performing rejection sampling to create a set of physically feasible trajectories Tadv and then projecting the trajectory generated by on to the physically feasible set, measured by L2 distance. Our search space is low-dimensional and conducive to query-based black box optimization, while still allowing for fine-grained actor motion control.”)).
Regarding claim 11, Wang discloses the invention of claim 6 as discussed above. Wang further discloses comprising rejecting a value of the first parameter as the modified value based on association of said value with a rejection criterion (see Wang pg. 9912, 2nd column (“To increase the perturbed trajectory’s plausibility, we ensure it does not collide with other actors or the original expert trajectory of the SDV. In practice, we do this by first performing rejection sampling to create a set of physically feasible trajectories Tadv and then projecting the trajectory generated by on to the physically feasible set, measured by L2 distance. Our search space is low-dimensional and conducive to query-based black box optimization, while still allowing for fine-grained actor motion control.”)).
Regarding claim 12, Wang discloses the invention of claim 6 as discussed above. Wang further discloses requesting the log data based on an estimated likelihood of a driving sequence associated with the log data occurring in the real-world environment (see Wang pg. 9914, 1st column (“Dataset We evaluate our approach on a self-driving dataset, UrbanScenarios, which has 5,000 driving logs of 25 seconds each. Our dataset is collected across multiple cities in North America, and contains different types of map layouts and varying traffic densities. We curate the dataset and select interesting candidate scenarios to apply AdvSim on, where the SDV in the original scenario “interacts” with other vehicles.”)).
Regarding claim 13, Wang discloses the invention of claim 6 as discussed above. Wang further discloses wherein the log data is measured by sensors of the vehicle traversing the real-world environment (see Wang pg. 9911, 2nd column (“Our objective is to generate realistic challenging scenarios that cause autonomy system failure. We frame our objective as a black box adversarial attack that exercises every component of the autonomy system, including object detection, motion forecasting and motion planning. As we search over the space of realistic perturbations in actor motions of an existing scenario, we must update the sensor data that the SDV observes and then evaluate the autonomy system.”)).
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or non-obviousness.
Claims 1-5 and 14-19 are rejected under 35 U.S.C. 103 as being unpatentable over Wang in view of Colgate (US 2021/0004017 A1; published Jan. 7, 2021).
Regarding claim 1, Wang discloses [a] system comprising: (see Wang pg. 9909 Abstract (“In this paper, we propose AdvSim, an adversarial framework to generate safety-critical scenarios for any LiDAR-based autonomy system. Given an initial traffic scenario, AdvSim modifies the actors’ trajectories in a physically plausible manner and updates the LiDAR sensor data to match the perturbed world. Importantly, by simulating directly from sensor data, we obtain adversarial scenarios that are safety-critical for the full autonomy stack.”)).
Wang does not expressly disclose:
one or more processors; and (but see Colgate ¶ 316 (“The example computer system 2400 includes a processor 2402 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), one or more application specific integrated circuits (ASICs), one or more radio-frequency integrated circuits (RFICs), or any combination of these), a main memory 2404, and a static memory 2406, which are configured to communicate with each other via a bus 2408.”))
one or more non-transitory computer-readable media storing computer-executable instructions that, when executed, cause the system to perform operations comprising: (but see Colgate ¶ 317 (“The storage unit 2416 includes a machine-readable medium 2422 on which is stored instructions 2424 (e.g., software) embodying any one or more of the methodologies or functions described herein. The instructions 2424 (e.g., software) may also reside, completely or at least partially, within the main memory 2404 or within the processor 2402 (e.g., within a processor's cache memory) during execution thereof by the computer system 2400, the main memory 2404 and the processor 2402 also constituting machine-readable media.”)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang to incorporate the teachings of Colgate to implement AdvSim on a processor and memory, at least because doing so would enable, given an initial traffic scenario, modifying the actors’ trajectories in a physically plausible manner and updates the LiDAR sensor data to match the perturbed world. See Wang Abstract.
Wang further discloses:
receiving log data associated with a vehicle traversing a real-world environment, the log data associated with a first value of a first parameter associated with the environment; (see Wang pg. 9910, 1st column (“In this paper, we leverage real world traffic scenarios available in standard self-driving datasets…”); see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature. Candidate adversary trajectories can be generated by perturbing the change of curvature ˙κt and acceleration values [e.g., first value of a first parameter] at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”); see Wang pg. 9914, Section 4.1 (“Dataset We evaluate our approach on a self-driving dataset, UrbanScenarios, which has 5,000 driving logs of 25 seconds each. Our dataset is collected across multiple cities in North America, and contains different types of map layouts and varying traffic densities.”))
determining a modified value of the first parameter associated with a first event, the determining comprising: (see Wang pg. 9910, 1st column (“…and optimize the actor’s trajectories jointly to increase the risk of an autonomy system failure.”); see Wang pg. 9912, 2nd column (“Candidate adversary trajectories can be generated by perturbing the change of curvature κ˙ t and acceleration values at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32]. Moreover, to enlarge the space of sampled adversarial behaviors, we also allow the perturbation of initial states (x0, y0, θ0, v0) within set bounds. In summary, the perturbation space can be depicted as δ = {Δs0, (a0, κ˙ t|t=0) , . . . , (aT−1, κ˙ t|t=T−1)}.”))
determining a first intermediate value of the first parameter; (see Wang pg. 9910, 1st column (“As our perturbation modifies the actor’s trajectories, we need to adjust the sensor data to accurately reflect the actors’ new locations [e.g., first intermediate value]. We therefore adopt a high-fidelity LiDAR simulator…that modifies the sensor data accordingly taking into account occlusions.”); see Wang pg. 9912, 2nd column (“Candidate adversary trajectories can be generated by perturbing the change of curvature κ˙ t and acceleration values at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”))
determining, based at least in part on one or more calculations or heuristics, a first predicted driving scenario based at least in part on the first intermediate value of the first parameter, the predicted driving scenario corresponding to a driving scenario represented in the log data; (see Wang pg. 9910, 1st column (“After running the black-box autonomy system [e.g., calculations or heuristics] with modified sensor data [e.g., first intermediate value] as input, we obtain the planned trajectory [e.g., first predicted driving scenario]1 and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors.”), pg. 9914 1st column (“)
calculating a first score associated with the first predicted driving scenario, the first score descriptive of proximity to occurrence of the first event in the first predicted driving scenario; (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors.”); see Wang pg. 9913 (“Since we aim for a general adversarial scenario generation framework, we consider the autonomy system as a black box, where we access the evaluation scores [e.g., first score] through limited queries. Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost.”))
determining a second intermediate value of the first parameter; (see Wang pg. 9913, 2nd column (“AdvSim is a framework that can use any black-box search algorithm to identify autonomy system failures. The search algorithm attempts to find the safety critical scenarios by maximizing the adversarial objective Ladv in Eq. 2. The search algorithm queries the autonomy system with a candidate perturbation τadv to obtain a query pair (τadv,Ladv) and maintains a history H of past query pairs to generate the next candidate perturbation [e.g., second intermediate value of the first parameter].”))
determining that the second intermediate value satisfies a simulation condition; (Wang pg. 9912 2nd column (“To increase the perturbed trajectory’s plausibility, we ensure it does not collide with other actors or the original expert trajectory of the SDV [e.g., simulation condition]. In practice, we do this by first performing rejection sampling to create a set of physically feasible trajectories Tadv and then projecting the trajectory generated by δ on to the physically feasible set, measured by L2 distance.”))
executing, in response to determining that the second intermediate value satisfies the simulation condition, a run of a driving simulation based at least in part on the second intermediate value of the first parameter, the driving simulation including a simulation of an object represented in the log data, wherein the simulation of the object is controlled using a planning component; (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors.”); see Wang pg. 9910, 2nd column (“Recently, interpretable neural motion planners…provide an alternative that inherits the advantages of traditional pipelines and end-to-end approaches, by maintaining modularity and interpretability while enabling end-to-end learning…Our work evaluates a wide range of autonomy systems, including modular and end-to-end interpretable ones.”))
calculating a second score associated with the run of the driving simulation, the second score descriptive of proximity to occurrence of the first event in the run of the driving simulation; and (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors.”); see Wang pg. 9913 (“Since we aim for a general adversarial scenario generation framework, we consider the autonomy system as a black box, where we access the evaluation scores through limited queries.”))
identifying the second intermediate value as the modified value of the first parameter; (see Wang pg. 9913, 1st column (“Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost.”))
generating modified log data from the log data based at least in part on the modified value of the first parameter; and (see Wang pg. 9910, 1st column (“As our perturbation modifies the actors’ trajectories, we need to adjust the sensor data to accurately reflect the actors’ new locations. We therefore adopt a high-fidelity LiDAR simulator [27] that modifies the sensor data accordingly taking into account occlusions.”))
training a machine learning model for controlling an autonomous vehicle based at least in part on the modified log data (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors. We demonstrate the flexibility and scalability of our approach by generating over 4000 adversarial scenarios for a wide range of modern autonomy systems. Finally, we leverage AdvSim-generated safety-critical scenarios in training and further improve the safety of autonomy systems.”)).
Regarding claim 2, Wang, in view of Colgate, discloses the invention of claim 1 as discussed above. Wang further discloses wherein the first parameter is associated with conversion of an object represented in the log data to being controlled by a planning component in the driving simulation (see Wang pg. 9910, 2nd column (“Recently, interpretable neural motion planners [48, 49, 36] provide an alternative that inherits the advantages of traditional pipelines and end-to-end approaches, by maintaining modularity and interpretability while enabling end-to-end learning. This first began with joint perception and prediction [25, 10], which neural planners extended to include planning. Specifically, NMP [48] shared feature representations between multiple subtasks and predicted a cost volume to represent the quality of possible locations in planning. DSDNet [49] proposed an energybased model to parameterize the joint distribution of the actors’ future trajectories. P3 [36] developed a semantic occupancy representation and generated consistent ego-vehicle plans.”); see Wang pg. 9911, 1st column (“We choose to represent the behavior of actors as kinematic bicycle-model trajectories, allowing for physical feasibility and fine-grained behavior control.”)).
Regarding claim 3, Wang, in view of Colgate, discloses the invention of claim 1 as discussed above. Wang further discloses the operations comprising identifying the first parameter as contributing to proximity to occurrence of the first event (see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature.”)).
Regarding claim 4, Wang, in view of Colgate, discloses the invention of claim 1 as discussed above. Wang further discloses the operations comprising determining that the second score meets a threshold condition, and/or determining that the second score represents a minimum score (see Wang pg. 9913, 1st column (“Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost.”);
Regarding claim 5, Wang, in view of Colgate, discloses the invention of claim 1 as discussed above. Wang further discloses the operations comprising:
determining a further modified value of the first parameter; (see Wang pg. 9913, 2nd column (“We summarize our proposed AdvSim framework in Algorithm 1. Given an initial traffic scene, we pick the actors to be perturbed using heuristics, such as the closest reachable actors, and then sample physically plausible trajectories Tadv to ensure that our perturbations remain in this set. We then obtain the perturbation (k) at iteration k based on historical observations H using a selected black-box search algorithm (L. 5). We roll out the kinematics bicycle model states with initial state s0 and the perturbation (k), and project onto the feasible set Tadv to obtain the adversarial trajectories for the perturbed actors (L. 6).”))
generating further modified log data from the log data, the modified log data based at least in part on the further modified value of the first parameter; and (see Wang pg. 9914, 1st column (“After that, we update the sensor data accordingly (L. 7) and evaluate the full autonomy system on generated scenarios to compute Ladv (L. 8-9). Finally, after running the procedure for N iterations, we obtain the adversarial behaviors of perturbed actors as well as corresponding simulated LiDAR data.”))
training the machine learning model based at least in part on the further modified log data (see Wang pg. 9915, 1st column (“We investigate whether the robustness of the autonomy systems can be improved with our generated scenarios. We test several training schemes. First, we propose a curriculum learning (CL) [4] baseline where we first train on standard examples till convergence (easy examples), and then train on real challenging scenarios selected based on reachable actors (Sec 4.1) (hard examples). Then, we propose a robust training approach to leverage simulated worst-case scenarios. Specifically, we use AdvSim to generate a large number of adversarial scenarios to augment the training data. As discussed in Sec 3.2, the original expert trajectory is still a valid planning solution to mimic in the new scenario with respect to collisions, as we impose constraints on the perturbation. This allows us to re-train autonomy systems with scenarios produced by AdvSim using the same expert trajectories as ground truth.”)).
Regarding claim 14, Wang discloses […] perform operations comprising:
receiving log data associated with a vehicle traversing a real-world environment, the log data associated with a first value of a first parameter associated with the environment; (see Wang pg. 9910, 1st column (“In this paper, we leverage real world traffic scenarios available in standard self-driving datasets…”); see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature. Candidate adversary trajectories can be generated by perturbing the change of curvature ˙κt and acceleration values [e.g., first value of a first parameter] at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”); see Wang pg. 9914, Section 4.1 (“Dataset We evaluate our approach on a self-driving dataset, UrbanScenarios, which has 5,000 driving logs of 25 seconds each. Our dataset is collected across multiple cities in North America, and contains different types of map layouts and varying traffic densities.”))
inputting the log data into a search algorithm; (see Wang pg. 9913, 2nd column (“AdvSim is a framework that can use any black-box search algorithm [e.g., search algorithm] to identify autonomy system failures. The search algorithm attempts to find the safety critical scenarios by maximizing the adversarial objective Ladv in Eq. 2. The search algorithm queries the autonomy system with a candidate perturbation τadv to obtain a query pair (τadv,Ladv) and maintains a history H of past query pairs to generate the next candidate perturbation. We study a wide variety of black-box search algorithms including (1) Bayesian optimization [40, 35] (BO), (2) genetic algorithms [2] (GA), (3) random search [17] (RS) and (4) gradient estimation methods (NES [19] and Bandit-TD [20]).”))
receiving, from the search algorithm, a first modified value of the first parameter, (see Wang pg. 9914, 1st column (“We then obtain the perturbation δ(k) at iteration k based on historical observations H using a selected black-box search algorithm (L.5). We rollout the kinematics bicycle model states with initial states 0 and the perturbation δ(k), and project on to the feasible set Tadv to obtain the adversarial trajectories for the perturbed actors (L.6).”) wherein the search algorithm determines the first modified value of the first parameter based on (i) a first score associated with a predicted driving scenario corresponding to the log data and (ii) a first simulation score, (see Wang pg. 9913, 2nd column (“AdvSim is a framework that can use any black-box search algorithm to identify autonomy system failures. The search algorithm attempts to find the safety critical scenarios by maximizing the adversarial objective Ladv [e.g., first score associated with a predicted driving scenario] in Eq. 2. The search algorithm queries the autonomy system with a candidate perturbation τadv to obtain a query pair (τadv,Ladv) and maintains a history H of past query pairs to generate the next candidate perturbation.”), pg. 9914 (“After that, we update the sensor data accordingly (L.7) and evaluate the full autonomy system [e.g., first simulation score] on generated scenarios to compute Ladv (L.8-9).”) the first simulation score associated with occurrence of a first event in a first run of a driving simulation, (see Wang pg. 9913, 2nd column (“To induce autonomy system failures, we propose a combination of three costs as our adversarial loss function. These costs are similar to those autonomy systems [37,52] attempt to minimize over in Eq. 1. We first include lIL, an imitation-learning based cost that encourages the SDV’s output plan to deviate from the recorded human trajectory in the original scenario. We compute this as a smooth ℓ1 distance between output trajectory τ∗0 and the ground-truth for the entire scenario horizon. We also compute a cumulative collision cost ltcol that encourages the perturbation to cause the SDV to collide with other actors in the scene. Finally, we add a safety cost cts (xadv,τ∗0) that encourages the output plan τ∗0 to have lane violations and be dangerous (i.e. high accelerations and jerk) at each timestep t.”)) the first run of the driving simulation including an autonomous vehicle and based at least in part on the first modified value of the first parameter (see Wang pg. 9913, 1st column (“Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost.”));
outputting first modified log data based at least in part on the first modified value of the first parameter; and (see Wang pg. 9910, 1st column (“As our perturbation modifies the actors’ trajectories, we need to adjust the sensor data to accurately reflect the actors’ new locations. We therefore adopt a high-fidelity LiDAR simulator [27] that modifies the sensor data accordingly taking into account occlusions.”))
training a machine learning model for controlling an autonomous vehicle based at least in part on the first modified log data (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors. We demonstrate the flexibility and scalability of our approach by generating over 4000 adversarial scenarios for a wide range of modern autonomy systems. Finally, we leverage AdvSim-generated safety-critical scenarios in training and further improve the safety of autonomy systems.”)).
Wang does not expressly disclose [o]ne or more non-transitory computer-readable media storing instructions executable by one or more processors, wherein the instructions, when executed, cause the one or more processors to (but see Colgate ¶ 317 (“The storage unit 2416 includes a machine-readable medium 2422 on which is stored instructions 2424 (e.g., software) embodying any one or more of the methodologies or functions described herein.”)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang to incorporate the teachings of Colgate to implement AdvSim on a machine-readable medium memory storing instructions embodying AdvSim, at least because doing so would enable, given an initial traffic scenario, modifying the actors’ trajectories in a physically plausible manner and updates the LiDAR sensor data to match the perturbed world. See Wang Abstract.
Regarding claim 15, Wang, in view of Colgate, discloses the invention of claim 14 as discussed above. Wang further discloses wherein the first event is associated with an overlap between a representation in the first run of the driving simulation of a first object in the environment and a representation in the first run of the driving simulation of a second object in the environment (see Wang pg. 9910, 1st column (“After running the black-box autonomy system with modified sensor data as input, we obtain the planned trajectory and evaluate how adversarial the scenario was. Our adversarial objective captures multiple safety factors such as collisions, violations in traffic rules, and uncomfortable driving behaviors.”)).
Regarding claim 16, Wang, in view of Colgate, discloses the invention of claim 14 as discussed above. Wang further discloses wherein the first parameter is associated with a geometric attribute associated with a first object in the environment, a speed associated with the first object, metadata associated with the first object, an environmental attribute associated with the environment; a map attribute associated with the environment, a temporal attribute associated with the environment; or a conversion of an object represented in the log data to being controlled by a planning component in the driving simulation (see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature.”); see Wang pg. 9914, Section 4.1 (“Dataset We evaluate our approach on a self-driving dataset, UrbanScenarios, which has 5,000 driving logs of 25 seconds each. Our dataset is collected across multiple cities in North America, and contains different types of map layouts and varying traffic densities.”)).
Regarding claim 17, Wang, in view of Colgate, discloses the invention of claim 14 as discussed above. Wang further discloses wherein the search performs a directed search to determine the first modified value of the first algorithm (see Wang pg. 9911 2nd column (“Our objective is to generate realistic challenging scenarios that cause autonomy system failure. We frame our objective as a black box adversarial attack that exercises every component of the autonomy system, including object detection, motion forecasting and motion planning. As we search over the space of realistic perturbations in actor motions of an existing scenario, we must update the sensor data that the SDV observes and then evaluate the autonomy system.”)).
Regarding claim 18, Wang, in view of Colgate, discloses the invention of claim 14 as discussed above. Wang further discloses wherein the log data comprises a second value of a second parameter associated with the environment, and wherein the operations comprise: (see Wang pg. 9912, 2nd column (“To produce physically feasible actor behaviors, we parameterize the trajectory τadv = {st}Tt =0 as a sequence of kinematic bicycle model states st = {xt, yt, θt, vt, κt, at} in the next T timesteps. Here (x, y) is the center position of the perturbed actor, θ is the heading, v and a are the forward velocity and acceleration, and κ is the vehicle path’s curvature. Candidate adversary trajectories can be generated by perturbing the change of curvature κ˙ t and acceleration values at within set bounds at different timesteps, and using the kinematic bicycle model to compute the other states [32].”))
receiving, from the search algorithm, a second modified value of the second parameter, wherein the search algorithm determines the second modified value of the second parameter based on a second simulation score, the second simulation score associated with occurrence of the first event in a second run of the driving simulation, the second run of the driving simulation based at least in part on the second modified value of the second parameter; and (see Wang pg. 9913, 1st column (“Since we aim for a general adversarial scenario generation framework, we consider the autonomy system as a black box, where we access the evaluation scores through limited queries. Our goal is to find the perturbation that maximizes the SDV’s planned trajectory cost. In this section, we introduce the adversarial objective we optimize to produce worst-case scenarios and detail the search algorithms applied. We then summarize the AdvSim algorithm.”); see Wang pg. 9913, 2nd column (“Adversarial Objective: To induce autonomy system failures, we propose a combination of three costs as our adversarial loss function. These costs are similar to those autonomy systems [37, 52] attempt to minimize over in Eq. 1. We first include lIL, an imitation-learning based cost that encourages the SDV’s output plan to deviate from the recorded human trajectory in the original scenario. We compute this as a smooth ℓ1 distance between output trajectory τ 0 and the ground-truth for the entire scenario horizon. We also compute a cumulative collision cost lt col that encourages the perturbation to cause the SDV to collide with other actors in the scene.”))
outputting second modified log data based at least in part on the second modified value of the first parameter (see Wang pg. 9913, 2nd column (“Our use of multiple different costs allows us to identify different types of autonomy system failures, such as unnatural trajectories, collisions, and hard braking.”)).
Regarding claim 19, Wang, in view of Colgate, discloses the invention of claim 18 as discussed above. Wang further discloses receiving, from the search algorithm, a set of modified values of the first parameter and the second parameter associated with occurrence of the first event (see Wang pg. 9912, 2nd column (“Moreover, to enlarge the space of sampled adversarial behaviors, we also allow the perturbation of initial states (x0, y0, θ0, v0) within set bounds. In summary, the perturbation space can be depicted as δ = {Δs0, (a0, κ˙ t|t=0) , . . . , (aT−1, κ˙ t|t=T−1)} .”); see Wang pg. 9913, 2nd column (“We summarize our proposed AdvSim framework in Algorithm 1. Given an initial traffic scene, we pick the actors to be perturbed using heuristics, such as the closest reachable actors, and then sample physically plausible trajectories Tadv to ensure that our perturbations remain in this set. We then obtain the perturbation (k) at iteration k based on historical observations H using a selected black-box search algorithm (L. 5).”)).
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 SHAHID KHAN whose telephone number is (571)270-0419. The examiner can normally be reached M-F, 9-5 est.
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, Usmaan Saeed can be reached at (571)272-4046. 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.
/SHAHID K KHAN/Primary Examiner, Art Unit 2146
1 “In some examples, determining a predicted driving scenario may comprise executing a run of a driving simulation.” Spec. ¶ 18.
Prosecution Insights