Object Motion Prediction Training With Non-Differentiable Priors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle systems struggle to accurately predict the motion of objects in complex environments due to the lack of incorporation of non-differentiable prior knowledge, such as lane geometry and traffic rules, into their motion forecasting models.
Innovation Solution
The system employs a machine-learned object motion prediction model that incorporates non-differentiable prior knowledge through a REINFORCE gradient estimation technique, allowing it to train the model using reward functions that encode prior knowledge about object motion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If non-differentiable prior knowledge (lane geometry, traffic rules) is incorporated into motion prediction models, then prediction accuracy and safety are improved, but computational complexity and training difficulty increase
Solution Approach 1:
The patent introduces an intermediary reward function that bridges the gap between non-differentiable prior knowledge and differentiable model training. The reward function encodes lane geometry constraints and traffic rules in a form that can guide gradient-based optimization without requiring the model to directly process non-differentiable constraints, thus improving prediction reliability while managing training complexity
Solution Approach 2:
The patent transforms the training objective by changing the parameter optimization approach. Instead of directly optimizing for constraint satisfaction (non-differentiable), the system optimizes for reward maximization where the reward is computed based on constraint adherence. This parameter transformation enables the use of gradient-based methods while incorporating prior knowledge about lane geometry and traffic rules
2Reliability
If REINFORCE gradient estimation technique is used to train the model with prior knowledge, then adherence to traffic rules is improved, but training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-encoding traffic rules and lane geometry into the reward function structure before training begins. This preprocessing of constraints into the reward formulation allows the REINFORCE algorithm to efficiently optimize adherence to these rules during training, rather than learning them from scratch, thereby reducing overall training time while maintaining high adherence levels
Data Source
AI summary
The present disclosure provides systems and methods for training probabilistic object motion prediction models using non-differentiable representations of prior knowledge. As one example, object motion prediction models can be used by autonomous vehicles to probabilistically predict the future location(s) of observed objects (e.g., other vehicles, bicyclists, pedestrians, etc.). For example, such models can output a probability distribution that provides a distribution of probabilities for the future location(s) of each object at one or more future times. Aspects of the present disclosure enable these models to be trained using non-differentiable prior knowledge about motion of objects within the autonomous vehicle's environment such as, for example, prior knowledge about lane or road geometry or topology and/or traffic information such as current traffic control states (e.g., traffic light status).