Vehicle Trajectory Training With Perturbation Recovery Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained on ideal or expert trajectories fail to accurately imitate real-world vehicle trajectories, leading to compounded errors and potential system failures due to unforeseen states, known as the covariate shift problem.
Innovation Solution
Introduce perturbations to vehicle states in expert trajectories, generate modified trajectories that converge to the original, and incorporate a collision loss function to penalize unsafe scenarios, ensuring the model accounts for deviations and maintains safety.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained using ideal or expert trajectories, then the model can learn optimal driving behavior, but the model fails to handle real-world deviations and accumulates errors leading to system failures
Solution Approach 1:
The patent applies preliminary action by proactively introducing perturbations to training trajectories before deployment. The system pre-modifies expert trajectories by adding noise and deviations to create synthetic training data that anticipates real-world variations, enabling the model to learn robust behavior in advance rather than failing when encountering unseen states during actual operation
Solution Approach 2:
The patent implements dynamics by transforming static expert trajectories into dynamic training samples through perturbation. The system continuously generates varied trajectory versions by applying different noise levels and types, making the training data adaptable to diverse real-world conditions rather than relying on fixed ideal paths
2Manufacturing precision
If the vehicle strictly follows expert trajectories, then optimal performance is achieved, but small errors compound over time causing the vehicle to reach unseen states
Solution Approach 1:
The patent applies beforehand cushioning by training the model on perturbed trajectories that include potential error states. This creates a buffer of learned resilience, where the model has already encountered and recovered from simulated deviations during training, preventing error compounding during actual deployment
3Adaptability or versatility
If perturbations are applied to training trajectories, then the model learns to handle deviations, but the training complexity and computational requirements increase
Solution Approach 1:
The patent applies parameter changes by systematically varying perturbation parameters (noise type, magnitude, frequency) to generate diverse training samples from a single expert trajectory. This approach increases adaptability without proportionally increasing complexity, as the same base trajectory can generate multiple perturbed versions through parameter modification rather than requiring entirely new trajectories
Data Source
AI summary
Systems and methods are provided for training a machine learning model to generate a planned trajectory for an ego vehicle that accounts for deviations from expert trajectories used to train the machine learning model. Examples include obtaining a first trajectory for a machine learning model, the first trajectory comprises a sequence of a plurality of vehicle states, and perturbing at least one of the plurality of vehicle states. Examples also include generating a second trajectory based on the at least one perturbed vehicle state and smoothening the second trajectory to correspond to the first trajectory. Examples further include training the machine learning model using the smoothened second trajectory to produce a planned trajectory for controlling a vehicle.