Navigating Autonomous Vehicles Based On Modulation Of A World Model Representing Traffic Entities

What is this patent about?

’346 is related to the field of autonomous vehicle navigation and, more specifically, to predicting the behavior of traffic entities (pedestrians, cyclists, other vehicles) to improve the safety and naturalness of autonomous driving. Existing autonomous systems often struggle to accurately predict the intentions of these traffic entities, leading to erratic vehicle behavior such as sudden stops or unnecessary delays. This patent addresses the problem of predicting the behavior of traffic entities to enable safer and more human-like autonomous navigation.

The underlying idea behind ’346 is to leverage machine learning models trained on human observations to predict the 'hidden context' or internal state (intentions, awareness) of traffic entities. Instead of relying solely on motion parameters, the system uses images of traffic entities as input to a model that has been trained to mimic how humans interpret the behavior of other road users. This allows the autonomous vehicle to anticipate actions based on a more nuanced understanding of the situation.

The claims of ’346 focus on an autonomous vehicle receiving sensor data, generating a point cloud representation of its surroundings, identifying traffic entities, and then, for each entity, determining motion parameters and a 'hidden context' using a machine learning model. The system then determines a region in the point cloud where the traffic entity is expected to be within a certain time, and modifies this region based on the predicted hidden context. Finally, the autonomous vehicle navigates to maintain a safe distance from these modified regions.

In practice, the system uses sensors like cameras and lidars to perceive the environment. The images captured are fed into a machine learning model that outputs summary statistics representing the predicted hidden context of each traffic entity. This hidden context is then used to adjust the predicted future location of the traffic entity in the point cloud. For example, if the model predicts that a pedestrian intends to cross the street, the region representing their potential future location is extended in the crossing direction, causing the autonomous vehicle to adjust its path to maintain a safe distance.

This approach differentiates itself from prior solutions by incorporating human-like reasoning into the autonomous navigation system. Instead of simply reacting to observed motion, the system anticipates behavior based on predicted intentions. Furthermore, the system can adjust its behavior based on the confidence level of the hidden context prediction. For example, if the model is uncertain about a pedestrian's intentions, the system may adopt a more conservative approach, increasing the safety margin to avoid potential collisions.