Obviousness

Obviousness Analysis under 35 U.S.C. § 103 for US Patent 11520346

This section analyzes the obviousness of US Patent 11520346 under 35 U.S.C. § 103, identifying potential combinations of prior art references and the motivation a person having ordinary skill in the art (POSITA) would have had to combine them. The analysis focuses on the independent claims (Claim 1, Claim 9, and Claim 16) as presented in the patent summary.

General Principles of Obviousness:

For a patent claim to be obvious, a POSITA at the time of the invention would have been motivated to combine existing prior art references, or modify a prior art reference, with a reasonable expectation of success. This motivation can stem from the knowledge of a POSITA, the prior art itself, or the nature of the problem to be solved. Generalized assertions of technological predictability or merely wanting to build something "better," "more efficient," or with "more features" are insufficient to establish motivation to combine.

Prior Art Landscape in Autonomous Vehicles (AVs):

The field of autonomous vehicles is characterized by extensive prior art in navigation, object detection, and path planning. Key areas include AI/ML algorithms, LiDAR sensors, vehicle-to-everything (V2X) communication, high-definition (HD) mapping, cybersecurity, and simulation technologies. Sensor fusion, combining data from multiple sensors like cameras, LiDAR, and radar, is crucial due to the limitations of individual sensors in varying conditions. Machine learning, particularly deep learning, is widely applied to sensor data fusion to improve object detection and decision-making by automatically learning features from large datasets. Autonomous vehicle systems also commonly employ "world models" for perception and trajectory prediction.

Analysis of Independent Claims:

Claim 1: Method for navigating an autonomous vehicle

Claim 1 encompasses:

1. Receiving sensor data.
2. Identifying traffic entities.
3. Determining motion parameters and "hidden context" using a machine learning model trained with human feedback on traffic scenarios.
4. Navigating based on motion parameters and hidden context.

Obviousness Argument:

A POSITA at the time of the invention (priority date January 30, 2019) would have found Claim 1 obvious by combining the teachings of conventional autonomous vehicle navigation systems (e.g., US10338594B2 or US20170090478A1) with the known application of machine learning for predicting human behavior and the common practice of using human feedback for training such models.

• Conventional AV Navigation (Elements 1, 2, 4): Autonomous vehicles routinely receive sensor data (e.g., camera images, LiDAR scans) to perceive their environment, identify traffic entities (pedestrians, other vehicles), determine their motion, and navigate accordingly to avoid collisions. [cite: The full patent text confirms this information] US10338594B2, for example, discusses autonomous driving based on sensor data, object recognition (vehicles or pedestrians), and determining vehicle control solutions. US20170090478A1 also describes systems for automatically navigating a vehicle using data from physical sensors, monitoring the vehicle environment, and determining control solutions. These references clearly disclose the receipt of sensor data, identification of traffic entities, and navigation based on motion.
• Machine Learning for Predicting Behavior (Element 3 - "Hidden Context" and ML Model): The use of machine learning algorithms to perceive the environment and make decisions is fundamental to autonomous vehicles. Predicting human behavior (e.g., intentions of pedestrians or other drivers) is a known challenge in autonomous driving. Prior art discusses the need for AVs to understand road users' intentions. For example, systems integrating multimodal sensor fusion with generative models to facilitate natural language interaction for AVs demonstrate an understanding of the need for "situational awareness" beyond simple object classification. The concept of predicting "state of mind" or "intention" of road users is acknowledged as crucial for autonomous driving.
• Training with Human Feedback: The practice of training machine learning models with human-in-the-loop feedback is well-established in the field of robotics and AV data labeling. Systems are designed where humans verify and correct machine predictions, indicating that using human responses to train models for understanding complex, nuanced behavior like "hidden context" would be a natural extension.

Motivation to Combine:

A POSITA would have been motivated to combine these elements to address the known limitations of conventional AV systems that fail to accurately predict complex human behavior, leading to "unnatural movement" such as sudden stops or unnecessary waiting. [cite: The full patent text confirms this information] The problem of accurately predicting the motion of non-stationary objects like pedestrians and bicyclists, particularly their intentions (e.g., whether a pedestrian will cross the street or remain stationary, or if a bicyclist will change lanes), was a recognized challenge. [cite: The full patent text confirms this information]

Therefore, a POSITA would have sought to improve the accuracy of traffic entity behavior prediction in autonomous vehicles. Combining established AV navigation techniques with machine learning models trained on human assessments of "hidden context" would be an obvious solution to achieve a more human-like and safer navigation experience, addressing the problem of unpredictable human actions. The motivation is to enhance safety and naturalness of autonomous vehicle operation by better anticipating the nuanced behaviors of human road users.

Claim 9: Method for braking an autonomous vehicle

Claim 9 outlines a method for braking involving:

1. Receiving sensor data.
2. Identifying traffic entities.
3. Determining a "hidden context" using a machine learning model (trained with human feedback) and obtaining a measure of its statistical distribution.
4. Determining an activation threshold for the braking system based on this statistical distribution.
5. Predicting collision likelihood within the threshold.
6. Activating the braking system if a collision is likely.

Obviousness Argument:

Claim 9 would have been obvious by combining the elements found in conventional autonomous braking systems with the application of machine learning for probabilistic risk assessment and the use of statistical distributions from human-trained models.

• Conventional Braking Systems (Elements 1, 2, 5, 6): Autonomous vehicles include active safety systems that predict and avoid collisions, often taking automatic action like braking. [cite: The full patent text confirms this information] This involves receiving sensor data, identifying objects, predicting travel paths, and activating braking systems to prevent impending collisions. [cite: The full patent text confirms this information] US10338594B2, for example, describes a vehicle control unit applying increments to a braking system for recovery actions.
• Machine Learning for "Hidden Context" and Statistical Distribution (Elements 3, 4): As discussed for Claim 1, using machine learning to predict "hidden context" based on human feedback is a logical extension of existing AV perception. The output of such a model naturally includes statistical distributions (e.g., mean, variance, kurtosis, skew) that represent the uncertainty or agreement in human assessment. [cite: The full patent text confirms this information] Machine learning algorithms are designed to reduce uncertainty and provide more reliable information from heterogeneous sensor data.
• Activation Threshold Based on Distribution: A POSITA would understand that the reliability or certainty of a prediction (e.g., of a pedestrian's intention) should influence safety-critical actions like braking. If the prediction of a "hidden context" (like intent to cross the street) has high uncertainty (e.g., a broad statistical distribution), a more conservative approach to braking (e.g., a lower activation threshold, meaning braking earlier) would be a logical safety measure. The patent itself states that "low activation thresholds if the model outputs indicate higher uncertainty in values of hidden context attributes" is used, implying an inverse relationship between certainty and threshold. [cite: The full patent text confirms this information] This is a direct application of risk assessment based on predictive model confidence.

Motivation to Combine:

The motivation here is to improve the safety and reliability of autonomous braking systems by incorporating a more nuanced understanding of human behavior, especially in uncertain situations. Conventional systems might trigger false positives or react too late due to a lack of understanding of underlying human intentions. By leveraging the statistical distribution of the "hidden context" from a human-trained ML model, a POSITA would be motivated to dynamically adjust the braking activation threshold to minimize false positives while maximizing safety. This addresses the problem of conventional systems failing to accurately predict motion of non-stationary objects, leading to undesirable sudden stops. [cite: The full patent text confirms this information] The goal is to make the braking system more adaptive and intelligent, similar to how human drivers factor in uncertainty about other road users' intentions.

Claim 16: Method for modifying a world model

Claim 16 describes a method for modifying a world model involving:

1. Generating a point cloud representation from sensor data.
2. Identifying traffic entities.
3. Determining motion parameters.
4. Predicting a "hidden context" using a machine learning model (trained with human feedback).
5. Determining a future region where each traffic entity is expected to be.
6. Modifying this region based on the predicted hidden context.
7. Navigating to stay a threshold distance from the modified region.

Obviousness Argument:

Claim 16 would have been obvious by combining known techniques for building and using world models and occupancy grids in autonomous vehicles with the application of machine learning to predict "hidden context" and dynamically adjust predicted occupancy or danger zones.

• World Models and Point Clouds (Elements 1, 2, 3, 5): Autonomous vehicles commonly build "world models" and use sensor data (e.g., LiDAR scans) to generate point cloud representations of their surroundings. [cite: The full patent text confirms this information] These models identify objects, track their motion, and predict their future positions, often using occupancy grids to represent the environment and potential obstacles. [cite: The full patent text confirms this information] Advanced world models in autonomous driving include image-based and occupancy-based models, and frameworks often predict future trajectories.
• Machine Learning for "Hidden Context" (Element 4): As established in the analysis of Claim 1, using machine learning, particularly models trained with human feedback, to predict human "hidden context" (intentions, awareness) is a recognized advancement for improving AV perception.
• Modifying Regions Based on Hidden Context (Elements 6, 7): The core of this claim involves adjusting the predicted "safe zone" or "region of expected movement" of a traffic entity based on its predicted "hidden context." For example, if a pedestrian's "hidden context" indicates a high intent to cross the street, their predicted future path or the avoidance zone around them would logically be expanded in that direction. The patent itself describes this: "If a determination is made that the hidden context indicates that the user represented by the traffic entity is likely to move in the direction having a component along the motion vector, the region is extended along the direction of the motion vector." [cite: The full patent text confirms this information] This is a direct application of the "hidden context" information to a world model for navigation. The modification of a region (e.g., by stretching or decreasing its size) based on behavioral predictions to ensure a safe distance is a logical consequence of having more sophisticated behavioral insights.

Motivation to Combine:

The motivation for a POSITA to combine these elements is to create a more robust and adaptive world model that better anticipates dynamic human behavior in complex traffic scenarios. Conventional world models might rely solely on observed motion parameters, leading to conservative or inefficient navigation if human intentions are not accurately captured. By integrating the "hidden context" derived from human-trained ML models, a POSITA would be motivated to dynamically modulate the representation of traffic entities within the world model. This allows the autonomous vehicle to navigate more smoothly and safely by accounting for the probabilistic likelihood of human actions, improving both efficiency and safety in traffic. The goal is to resolve the inability of conventional techniques to accurately predict motion of non-stationary objects by introducing a "stable signal of potential behavior" to modulate the vehicle's performance envelope. [cite: The full patent text confirms this information]

Conclusion on Obviousness:

Based on the above analysis, the independent claims of US Patent 11520346 appear to be obvious under 35 U.S.C. § 103. The core innovation of using "hidden context" derived from human-trained machine learning models to influence autonomous vehicle navigation, braking, and world model modulation, while novel in its specific implementation, represents an incremental improvement on well-established principles in autonomous vehicle technology and artificial intelligence. The motivation to combine these existing concepts arises directly from the recognized challenges in making autonomous vehicles safer and more capable of handling complex, human-driven traffic environments in a natural manner.