Behavior Prediction Neural Networks Without Non-Causal Agent Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in predicting the future trajectories of agents due to the inability of neural networks to generalize across diverse scenarios, particularly when non-causal agents, such as parked cars, affect prediction accuracy, and collecting labeled data for robustness is expensive and difficult.
Innovation Solution
The system trains a behavior prediction neural network by perturbing training data to remove or modify non-causal agents, and evaluates its robustness and sensitivity to ensure accurate trajectory predictions across various environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural networks are trained on diverse scenarios including non-causal agents, then prediction coverage improves, but prediction accuracy deteriorates due to spurious correlations
Solution Approach 1:
The patent extracts and removes non-causal agents from training data through perturbation techniques. By systematically removing agents that do not causally influence target trajectories, the training process learns to ignore spurious features while maintaining sensitivity to causal factors, thereby improving prediction accuracy without sacrificing coverage
Solution Approach 2:
The patent applies preliminary perturbation to training data before model training. By pre-processing training examples to remove or modify non-causal agents, the model is trained in advance to be robust against such distractions, enabling accurate predictions across diverse scenarios without learning spurious correlations
2Reliability
If labeled data is collected to improve robustness evaluation, then reliability improves, but cost and difficulty increase
Solution Approach 1:
The patent enables the system to self-evaluate robustness by using the trained model to predict trajectories on perturbed versions of its own training data. By comparing predictions on original versus perturbed inputs, the system automatically identifies robustness issues without requiring external labeled test data, thereby reducing collection costs while maintaining evaluation reliability
Solution Approach 2:
The patent implements a feedback mechanism where prediction results on perturbed data feed back into robustness evaluation. The system uses the difference between predictions on original and perturbed training examples as a self-generated feedback signal to assess and improve robustness, eliminating the need for expensive external labeled datasets
Data Source
AI summary
Methods, systems, and apparatus for predicting future trajectories of agents in an environment. In one aspect, a system comprises one or more computers configured to receive a data set comprising multiple training examples. The training examples include scene data comprising respective agent data for multiple agents and a ground truth trajectory for a target agent that represents ground truth motion of the target agent after a corresponding time point. The one or more computers obtain data identifying one or more of the multiple agents as non-causal agents for each training example. A non-causal agent is an agent whose states do not cause the ground truth trajectory for the target agent to change. The one or more computers generate a respective modified training example from each of the multiple training examples.