Trajectory Prediction Oracle Unit for Collision Avoidance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated vehicles face challenges in accurately predicting the trajectories of other road users, particularly vulnerable road users like pedestrians and cyclists, due to their dynamic movements and disregard for traffic rules, which complicates collision avoidance and trajectory planning.
Innovation Solution
A system and method for training a generator unit and a discriminator unit using generative adversarial networks (GANs) with an oracle unit to predict future trajectories, where the oracle unit determines collision-free trajectories and influences the generator unit through reinforcement learning, utilizing an off-policy actor-critic framework and encoding static and dynamic environment information in polar angle space to minimize collisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional trajectory prediction methods are used, then the system is simpler to implement, but the prediction accuracy of vulnerable road user trajectories deteriorates due to their dynamic and unpredictable behavior
Solution Approach 1:
The patent introduces an oracle unit as an intermediary component that evaluates collision-free trajectories and provides guidance signals to the generator unit. This mediator enables the system to learn collision avoidance while maintaining the GAN architecture, resolving the contradiction between prediction accuracy and system complexity by adding a focused evaluation component rather than redesigning the entire system.
Solution Approach 2:
The patent implements feedback mechanisms where the discriminator unit and oracle unit provide evaluation signals back to the generator unit during training. The discriminator distinguishes between real and generated trajectories, while the oracle provides collision-free trajectory guidance, enabling the generator to iteratively improve prediction accuracy through feedback-driven learning.
2Adaptability or versatility
If the generator unit predicts all possible trajectories, then trajectory coverage is improved, but the ability to identify collision-free trajectories deteriorates without additional constraints
Solution Approach 1:
The oracle unit serves as a mediator that evaluates the reliability of generated trajectories by assessing collision-free properties. It provides guidance signals to the generator unit, enabling the system to maintain broad trajectory coverage while reliably identifying collision-free paths through the oracle's evaluation and feedback mechanism.
Solution Approach 2:
The patent replaces traditional mechanical collision avoidance constraints with a learning-based oracle unit that uses reinforcement learning to evaluate and guide trajectory generation. This substitution allows the system to explore diverse trajectories while learning collision avoidance patterns, improving both coverage and reliability.
3Measurement precision
If the system uses more training data to improve prediction accuracy, then the generator unit becomes more robust, but the training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-processing training data into observed trajectories and using the discriminator unit to pre-evaluate data quality during training. The oracle unit also provides preliminary guidance on collision-free trajectory patterns, enabling the system to learn from data more efficiently and reduce overall training time while maintaining accuracy.
Data Source
Figure 1
Figure 2
AI summary
An aspect of the invention describes a system for training a generator unit and a discriminator unit simultaneously, with said generator unit configured to determine a future trajectory of at least one other road in the environment of a vehicle considering an observed trajectory of the at least one other road user, with said discriminator unit configured to determine whether the determined future trajectory of the other road user is an actual future trajectory of the other road user, and with said system configured to train said generator unit and said discriminator unit simultaneously with gradient descent.