Autonomous Vehicle Motion Planning for Rare Event Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in being pre-trained to handle every possible scenario due to the limitless range of potential events, with rare 'corner cases' often not being captured in training data, making it inefficient to train on all possible situations, especially those that are extremely unlikely.
Innovation Solution
A system that identifies and generates simulated scenarios of safety threshold violations and near-collision events, altering actor trajectories and introducing disturbances to create realistic training data for an autonomous vehicle's motion planning model, focusing on likely corner cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the vehicle's motion planning model is trained on every possible scenario, then the model's ability to handle corner cases improves, but the training efficiency and practicality deteriorate due to the limitless range of possible scenarios
Solution Approach 1:
The system extracts and focuses only on the most critical and representative corner cases from the limitless scenario space. By identifying and isolating high-impact corner cases that are most likely to be encountered or most dangerous, the system trains the model on a curated subset rather than attempting to cover all possible scenarios, thus maintaining reliability while improving training efficiency
Solution Approach 2:
The system changes the parameters of training data selection by using importance sampling and risk-based filtering to identify corner cases. Instead of uniform sampling across all scenarios, the system adjusts selection parameters to prioritize scenarios with higher safety thresholds violations or near-collision risks, making the training process more practical and efficient
2Reliability
If the vehicle's perception system captures more rare events, then the training data quality improves, but the likelihood of capturing such events deteriorates due to their infrequency
Solution Approach 1:
The system creates synthetic copies of rare corner case events through simulation. By generating multiple simulated instances of rare events based on a small number of real captured events, the system enriches the training data quality without relying on the unlikely capture of numerous real rare events, thus resolving the contradiction between data quality and event quantity
Solution Approach 2:
The system performs preliminary simulation and analysis to identify which rare events are most critical before actual training. By pre-identifying high-value corner cases through simulation and importance sampling, the system prepares targeted training data in advance, making the most of limited captured rare events and improving training data quality without needing large quantities of captured events
Data Source
AI summary
Methods of identifying corner case simulation scenarios that are used to train an autonomous vehicle motion planning model are disclosed. A system selects a scene that includes data captured by one or more vehicles over a time period. The data includes one or more actors that the vehicle's sensors perceived over the time period in a real-world environment. The system selects a scene that includes a safety threshold violation, and it identifies the trajectory of an actor that participated in the violation. The system generates simulated scenes that alter the trajectory of the actor in the selected scene, selects simulated scenes that are more likely to occur in the real world and that may include safety threshold violations that go beyond any that may be found in the original scene, and uses the selected simulated scenes to train an autonomous vehicle motion planning model.