Autonomous Trajectory Model Training With Rasterized GAN Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine-learned models for predicting object trajectories in autonomous vehicles lack accuracy and efficiency, as they struggle to convert disparate data sets into comparable representations for effective training, leading to suboptimal prediction performance.
Innovation Solution
A computing system employing a generative-adversarial network architecture that converts predicted trajectories into rasterized trajectories spatially corresponding to rasterized images, using a differential rasterizer and discriminator model to determine loss scores and minimize overall loss, thereby improving the training efficiency and accuracy of prediction models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine-learned models are used for trajectory prediction, then the system can process data, but the prediction accuracy is insufficient due to inability to effectively convert disparate data sets into comparable representations
Solution Approach 1:
The patent introduces a discriminator model as an intermediary component that receives both rasterized images and predicted trajectories, converting them into a common representation space through loss score calculation. This mediator enables accurate comparison between disparate data types (images and trajectories) by projecting them into the same feature space, thereby resolving the accuracy issue without requiring complex manual feature engineering
Solution Approach 2:
The patent replaces traditional mechanical data processing approaches (manual feature extraction and alignment) with a learning-based system. The discriminator model automatically learns to convert disparate representations into comparable forms through adaptive learning rates and gradient-based optimization, eliminating the need for predefined processing pipelines and significantly improving prediction accuracy
2Measurement precision
If more complex training methods are employed to improve accuracy, then prediction performance improves, but training time and computational resources increase
Solution Approach 1:
The patent implements dynamic training through adaptive learning rates that automatically adjust during the training process. The discriminator and generator models use different learning rates that evolve over time, allowing the system to converge faster while maintaining high accuracy. This dynamic approach avoids the need for excessively long fixed-duration training sessions
Solution Approach 2:
The discriminator model provides continuous feedback to the generator model through loss score calculations. This feedback mechanism guides the generator to improve trajectory predictions iteratively, achieving high accuracy efficiently by focusing computational resources on the most informative training iterations rather than uniform processing
3Adaptability or versatility
If rasterized trajectories are converted to spatially correspond with rasterized images, then data comparability improves, but additional processing steps are required
Solution Approach 1:
The discriminator model serves multiple functions simultaneously: it converts rasterized images and predicted trajectories into a common representation space, calculates loss scores for training, and provides feedback for optimization. This multi-functional component achieves data compatibility without requiring separate processing modules for each transformation step, thereby managing complexity efficiently
Data Source
AI summary
Systems and methods for training machine-learned models are provided. A method can include receiving a rasterized image associated with a training object and generating a predicted trajectory of the training object by inputting the rasterized image into a first machine-learned model. The method can include converting the predicted trajectory into a rasterized trajectory that spatially corresponds to the rasterized image. The method can include utilizing a second machine-learned model to determine an accuracy of the predicted trajectory based on the rasterized trajectory. The method can include determining an overall loss for the first machine-learned model based on the accuracy of the predictive trajectory as determined by the second machine-learned model. The method can include training the first machine-learned model by minimizing the overall loss for the first machine-learned model.