Autonomous Trajectory Selection With Uncertainty-Triggered Fallback Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle systems lack the ability to accurately estimate the certainty of their outputs and are not robust to perturbations in their input space, making them unreliable for safety-critical applications due to the risk of failing without warning when encountering novel scenarios.
Innovation Solution
A hybrid architecture combining deep learning and classical methods for deterministic trajectory selection, incorporating uncertainty estimation through an out-of-distribution detector, fallback motion planner, and emergency planner to ensure safe operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If data-driven approaches and deep neural networks are used to replicate human behavior, then naturalistic driving trajectories are achieved, but the system lacks accurate uncertainty estimation and robustness to input perturbations
Solution Approach 1:
The patent combines data-driven deep neural networks with classical uncertainty estimation methods (such as Bayesian inference and ensemble methods) to create a hybrid system. This merging allows the system to maintain the naturalistic trajectory generation capabilities of deep learning while incorporating robust uncertainty estimation from classical methods, thereby resolving the contradiction between operational ease and reliability
Solution Approach 2:
The patent introduces an intermediary uncertainty estimation module that sits between the deep neural network and the trajectory execution system. This intermediary layer processes the network outputs and provides uncertainty quantification, acting as a mediator that enables both naturalistic behavior replication and reliable uncertainty assessment without requiring a complete system redesign
2Adaptability or versatility
If deep learning methods are used to achieve human-like trajectories, then adaptability to complex scenarios is improved, but the system fails without warning when encountering novel scenes outside training distribution
Solution Approach 1:
The patent performs preliminary uncertainty estimation and out-of-distribution detection before the autonomous vehicle executes trajectories or makes critical decisions. By proactively assessing uncertainty levels and comparing incoming data against training distribution characteristics, the system can prepare fallback responses in advance, preventing catastrophic failures without warning when encountering novel scenarios
Solution Approach 2:
The patent implements feedback mechanisms that continuously monitor the confidence and uncertainty of deep learning predictions. When uncertainty exceeds predefined thresholds or when out-of-distribution conditions are detected, the system triggers feedback loops that activate classical fallback methods or request human intervention, ensuring reliable failure detection while maintaining adaptability to complex scenarios
3Reliability
If classical methods like Kalman Filters are used, then uncertainty estimation is accurate, but the system lacks the robustness and generalization capability of deep learning approaches
Solution Approach 1:
The patent merges classical uncertainty estimation methods (such as Kalman Filters and Bayesian inference) with deep learning-based perception and prediction modules. This combination allows the system to leverage the mathematically rigorous uncertainty quantification of classical methods while maintaining the robustness and generalization capabilities of deep learning, particularly in handling perturbations and novel scenarios
Solution Approach 2:
The patent creates a composite system architecture that integrates multiple methodological 'materials' - deep neural networks for feature extraction and pattern recognition, classical statistical methods for uncertainty quantification, and rule-based systems for safety-critical decisions. This composite approach achieves both accurate uncertainty estimation and robustness to various input perturbations that neither method could achieve alone
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A system for deterministic trajectory selection based on uncertainty estimation includes a set of one or more computing systems. A method for deterministic trajectory selection includes receiving a set of inputs; determining a set of outputs; determining uncertainty parameters associated with any or all of the set of inputs and/or any or all of the set of outputs; and evaluating the uncertainty parameters and optionally triggering a process and/or action in response.