Autonomous Driving Controller Anomaly Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicle driving systems face challenges in accurately predicting the behavior of proximate objects, especially in anomalous situations where trained behavior models are insufficient, leading to potential dangerous outcomes.
Innovation Solution
Implementing a planning-based approach that generates a driving plan for proximate objects by optimizing a cost function from their perspective, allowing the vehicle to avoid collisions without relying on trained behavior models, and combining this with a learning-based approach for enhanced performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a learning-based approach using trained behavior models is used to predict proximate object behavior, then the system performs well under normal driving conditions, but it fails when confronted with anomalous situations not well-represented in the training data
Solution Approach 1:
The system segments the prediction approach into two distinct components: a learning-based prediction module that handles normal situations using trained behavior models, and a planning-based prediction module that handles anomalous situations by generating plans from first principles. This segmentation allows each module to specialize in its strengths while working together to cover the full range of driving scenarios.
Solution Approach 2:
The system introduces an intermediary anomaly detection mechanism that identifies when situations fall outside the training data distribution. This intermediary component acts as a bridge, determining when to switch from the learning-based approach to the planning-based approach, ensuring smooth transitions between different prediction modes based on situation normality.
2Productivity
If the system relies on trained behavior models for prediction, then it can efficiently handle common driving scenarios, but it produces poor or inaccurate predictions when faced with situations not well-represented by the trained model
Solution Approach 1:
The system dynamically adjusts its prediction strategy based on the characteristics of the current situation. When situations are normal and well-represented in training data, the efficient learning-based model is used. When anomalies are detected, the system dynamically switches to the planning-based approach, which sacrifices some computational efficiency but ensures accurate predictions in novel situations.
Solution Approach 2:
The system changes the fundamental parameters of the prediction approach based on situation type. For normal situations, it uses learned parameters from training data. For anomalous situations, it switches to generating plans based on physical constraints and cost functions, fundamentally changing how predictions are made to adapt to novel scenarios.
3Reliability
If the autonomous vehicle generates driving plans for proximate objects by optimizing cost functions, then it achieves more accurate collision avoidance in anomalous situations, but it increases computational complexity
Solution Approach 1:
The system applies the computationally intensive planning-based approach partially, only when anomalies are detected. For the majority of normal situations, it uses the more efficient learning-based model. This partial application of the complex planning algorithm reduces overall computational burden while maintaining high collision avoidance accuracy when it matters most.
Solution Approach 2:
The system performs preliminary anomaly detection before committing to the computationally intensive planning-based prediction. By pre-screening situations to identify only those that require planning-based approaches, the system avoids unnecessary computational complexity in normal situations while ensuring thorough analysis when anomalies are present.
Data Source
AI summary
Autonomous driving systems and methods include a controller of a first vehicle configured to detect an operating state of a second vehicle proximate the first vehicle, and to predict, based on the operating state of the second vehicle, a potential behavior for the second vehicle that optimizes a cost function from the perspective of the second vehicle. The controller then controls the first vehicle to avoid a collision with the second vehicle assuming the second vehicle operates according to the potential behavior.