Autonomous Vehicle Simulation Switching for Play-Forward Testing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for testing autonomous vehicle planning and control algorithms in offline simulations face uncertainty due to the use of sensor models that approximate actual perception systems, and playback of recorded data limits the system's ability to respond to divergent algorithm outputs.
Innovation Solution
A simulation system that identifies an event to be analyzed, receives data streams including perception, pose, and control information from vehicle event logs and a vehicle dynamics model, and switches between pure log execution and play-forward execution stages to simulate autonomous vehicle behavior, allowing divergence from original paths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If sensor models are used to generate surrogate perception data in simulation, then testing can be performed offline in reproducible simulated conditions, but the accuracy and reliability of test results are reduced due to approximation errors
Solution Approach 1:
The patent uses actual recorded perception data from real autonomous vehicle operation as a copy of real-world sensor output, replacing simulated surrogate data. This copying approach preserves the authentic characteristics and noise profiles of actual sensors while enabling offline testing through data playback, thereby maintaining both ease of testing and data accuracy.
2Measurement precision
If recorded log data is played back into algorithms under test, then actual perception data can be used for testing, but the system cannot respond to divergent algorithm outputs that differ from the original path
Solution Approach 1:
The patent segments the testing process into distinct phases: an initial phase where recorded log data is played back to establish baseline behavior, and a subsequent phase where the simulation diverges to test alternative algorithm outputs. This segmentation allows the system to maintain accuracy through real data while gaining adaptability to test different algorithmic decisions.
Solution Approach 2:
The patent introduces dynamic switching between data sources during simulation. The system transitions from playing back recorded log data to using generated surrogate data when the algorithm's output diverges from the original path. This dynamic approach enables the simulation to adapt its data source based on algorithm behavior, maintaining both accuracy and versatility.
3Stability of the object's composition
If the simulation follows the original recorded path strictly, then data consistency is maintained, but the ability to evaluate divergent algorithm behavior is lost
Solution Approach 1:
The patent introduces a hybrid data generation mechanism as an intermediary between strict log playback and completely synthetic simulation. This intermediary system generates surrogate perception data that is consistent with the original log data characteristics but allows for divergence when algorithm outputs differ from the original path. The intermediary maintains data consistency through learned patterns while enabling versatility through controlled divergence.
Data Source
AI summary
A system of creating a simulation to simulate behavior of an autonomous vehicle includes a simulation system having an electronic device and a computer-readable storage medium having one or more programming instructions. When executed, the one or more programming instructions cause the electronic device to identify an event that is to be analyzed, receive from an autonomous vehicle system a first data stream that includes event information from one or more vehicle event log files that corresponds to the event, receive from a vehicle dynamics model a second data stream that includes synthetic event information that corresponds to the event, until a switch point is detected operate in a pure log execution stage, upon detection of the switch point operate in a play-forward execution stage, and cause the new simulation to be executed.