Explainability System for Unanticipated Autonomous Vehicle Actions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles lack the ability to provide real-time explanations for unanticipated actions to passengers and other agents, leading to reduced confidence and potential misunderstandings during unexpected events.
Innovation Solution
An explainability system separate from the planning system generates real-time notifications that explain the reasoning behind unanticipated actions, using user profiles and historical data to customize notifications and optimize bandwidth and resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the autonomous vehicle provides real-time explanations for unanticipated actions, then passenger confidence and transparency are enhanced, but computational resources and bandwidth are consumed
Solution Approach 1:
The system pre-loads and stores historical driving data, user profiles, and contextual information in an expectation database before unanticipated events occur. This preliminary preparation enables the explainability system to quickly generate notifications without performing heavy computations in real-time, thus enhancing passenger confidence while minimizing computational resource consumption during critical moments.
Solution Approach 2:
The patent introduces a separate explainability system that acts as an intermediary between the autonomous vehicle's planning system and the passenger. This intermediary system processes and generates human-readable explanations, transforming complex computational decisions into understandable narratives without burdening the primary vehicle control systems, thereby preserving computational resources while improving transparency.
2Reliability
If the autonomous vehicle generates notifications for unanticipated actions, then transparency and confidence are improved, but latency increases
Solution Approach 1:
The system pre-processes and stores relevant contextual data, historical driving patterns, and user preference information in an expectation database before unanticipated events occur. This preliminary preparation allows the notification system to quickly retrieve and generate explanations without performing heavy data processing during the critical notification window, thus reducing latency while maintaining transparency.
Solution Approach 2:
The system generates notifications selectively based on the significance of unanticipated events and user preferences stored in the expectation database. Rather than notifying for every possible event, the system applies partial action by filtering and prioritizing notifications, thereby reducing overall latency and bandwidth consumption while still providing necessary transparency for important events.
3Ease of operation
If the system uses user profiles and historical data to customize notifications, then notification effectiveness is improved, but data processing complexity increases
Solution Approach 1:
The system pre-processes user feedback, preferences, and historical driving data during non-critical periods and stores processed results in the expectation database. This preliminary action transforms raw data into structured, easily queryable formats that facilitate customized notification generation without requiring complex real-time data processing, thus improving notification effectiveness while managing data processing complexity.
Data Source
AI summary
Provided are methods for generating notifications indicative of unanticipated actions. Data associated with a current trajectory of an autonomous vehicle, at least one constraint, data associated with historical data for a user using the autonomous vehicle from an expectation database, and data associated with a context that represents a relationship between the current trajectory, at least one constraint, and historical driving data is received. A model is deployed to determine, in real-time and based on the current trajectory, the at least one constraint, and the data associated with the context that a particular action by the autonomous vehicle is classified as an unanticipated autonomous vehicle action. The current trajectory, the at least one constraint, and the data associated with the context are analyzed within a predetermined range of time including a timestamp of the unanticipated autonomous vehicle action to determine a reason for occurrence of the unanticipated autonomous vehicle action. A notification is generated that includes the reason, wherein an intensity of the notification is based on, at least in part, the deviation.