Autonomous Vehicle Object Avoidance via Physics-Based State Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object avoidance in autonomous navigation rely heavily on accurate identification and classification of objects, which can be prone to errors and vulnerabilities from adversarial attacks, especially in scenarios with limited training data or obscured objects.
Innovation Solution
An autonomous vehicle employs worst-case motion assumptions for generic objects to predict future state boundaries, allowing it to estimate a maximum accessible area and adjust its path to avoid collisions, reducing the need for precise object identification and enhancing robustness against adversarial conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If accurate object identification and classification methods are used, then object avoidance accuracy is improved, but system reliability deteriorates due to vulnerabilities from adversarial attacks and errors in scenarios with limited training data or obscured objects
Solution Approach 1:
The patent extracts and removes the vulnerable object identification and classification components from the safety-critical path. Instead of relying on AI models to identify and classify objects, the system uses sensor data (LIDAR, radar, cameras) to directly measure physical parameters like distance, speed, and trajectory, and applies physics-based models to predict future states. This extraction eliminates the attack surface of neural networks while preserving the ability to detect and respond to objects.
Solution Approach 2:
The patent replaces the software-based AI identification system with a physics-based mechanical model. The system uses sensor measurements combined with physics equations (motion equations, collision dynamics) to predict object behavior and calculate safe navigation paths. This substitution replaces the vulnerable neural network with a deterministic physical model that is immune to adversarial attacks.
2Measurement precision
If object classification is performed to improve navigation accuracy, then navigation precision is improved, but device complexity increases due to the need for multiple sensors and processing systems
Solution Approach 1:
The patent makes the sensor system universal by using the same LIDAR, radar, and camera sensors for multiple purposes: detecting object presence, measuring distance and speed, tracking trajectory, and providing input to physics-based prediction models. This multi-functionality eliminates the need for separate classification systems while maintaining navigation precision.
Solution Approach 2:
The patent changes the parameters being measured and processed from categorical labels (object classes) to continuous physical parameters (distance, speed, acceleration, trajectory). This parameter transformation simplifies the system by working directly with sensor outputs that feed into physics equations, avoiding the complexity of classification algorithms while improving precision through direct physical measurement.
3Reliability
If worst-case motion assumptions are used to predict future state boundaries, then safety is improved, but loss of time increases due to conservative navigation decisions
Solution Approach 1:
The patent applies dynamics by continuously updating the physics-based prediction models with real-time sensor data. Instead of using static worst-case assumptions, the system dynamically adjusts object trajectory predictions based on measured position, speed, and acceleration. This allows the vehicle to navigate more efficiently by responding to actual object behavior while maintaining safety through continuous recalculation of future state boundaries.
Solution Approach 2:
The patent implements feedback by using sensor measurements to continuously verify and update the physics-based predictions of object motion. The system compares predicted trajectories with actual sensor data and adjusts its navigation decisions accordingly. This feedback loop enables the vehicle to maintain safety without excessive conservatism, reducing unnecessary delays while ensuring collision avoidance.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A method for autonomous navigation of an autonomous vehicle includes: estimating a stopping duration, for the autonomous vehicle to reach a full stop, based on a current speed of the autonomous vehicle; calculating a critical time from the current time by the stopping duration; detecting an object in a scan image, of a field proximal the autonomous vehicle, captured by a sensor on the autonomous vehicle at the current time; based on the scan image, deriving a current location and motion of the object; calculating a future state boundary that represents a ground area accessible to the object up to the critical time based on the current location and motion of the object and a set of predefined motion limit assumptions for generic objects proximal public roads; and electing a navigational action to avoid entry into the future state boundary prior to the critical time.