Autonomous Vehicle Planning Stack Spoofing for Yield-Assert Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles often struggle with making appropriate decisions when interacting with other entities on the road, such as yielding or asserting, leading to situations where they may be stopped behind slow-moving vehicles or fail to properly yield for pedestrians and bicycles.
Innovation Solution
A system that uses a machine learning model to evaluate interactions with agents, incorporating data from sensors and on-road experiences to determine whether to yield or assert, and employs an autolabeler module to generate and refine labels for training and spoofing the planning stack of the autonomous vehicle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If an autonomous vehicle uses a planning stack to make decisions about yielding or asserting, then the vehicle can avoid collisions, but it may be stopped behind slow-moving vehicles or fail to properly yield for pedestrians and bicycles
Solution Approach 1:
The system performs preliminary labeling of training data using the autolabeler module before actual model training, pre-processing the data to identify yield/assert patterns. This preliminary action enables the planning stack to learn from pre-processed, high-quality labeled data, improving decision-making effectiveness while maintaining collision avoidance capabilities
Solution Approach 2:
The system uses spoofing to create synthetic training data that copies real-world yield/assert scenarios. By generating artificial training examples that replicate actual driving situations, the system can train the planning stack without requiring extensive real-world data collection, thereby improving productivity while maintaining reliability through realistic scenario replication
2Reliability
If an autonomous vehicle avoids objects simply by detecting them with sensors, then collision avoidance is achieved, but the vehicle lacks sufficient safety for decisions beyond avoiding accidents
Solution Approach 1:
The system introduces a machine learning model as an intermediary between sensor detection and planning stack decisions. This intermediary layer processes sensor data and provides contextual understanding of yield/assert scenarios, enabling the vehicle to make nuanced decisions beyond simple collision avoidance while maintaining the reliability of object detection
Solution Approach 2:
The system performs preliminary labeling of training data using the autolabeler module to pre-process and annotate yield/assert scenarios. This preliminary action creates a structured knowledge base that the planning stack can use to make adaptive decisions, enhancing versatility while maintaining the reliability of basic collision avoidance through pre-processed training data
Data Source
AI summary
Disclosed herein are systems and method including a method for managing an autonomous vehicle. The method includes obtaining labels associated with various aspects of right-of-way interactions between an autonomous vehicle and an agent, wherein the right-of-way interactions occur when a human driver takes over for the autonomous vehicle and performs the right-of-way interactions, running an autonomous vehicle stack that is untrained for processing right-of-way interactions between the autonomous vehicle and the agent, injecting the labels into the autonomous vehicle stack and determining, based on the injecting of the labels into the autonomous vehicle stack, a performance of the autonomous vehicle stack.