Neural Trajectory Generation Using Weighted Basis Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating trajectories for autonomous vehicles using neural networks can produce unpredictable outputs, posing safety risks due to their 'black box' nature and lack of explainability.
Innovation Solution
A method that uses a neural network to generate trajectories by determining a set of basis trajectory elements based on map data and applying weights to these elements to constrain the output to reasonable and safe paths, ensuring that extreme outputs are avoided.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a neural network is used to generate trajectories directly, then the system can learn complex driving patterns from data, but the output becomes unpredictable and unsafe
Solution Approach 1:
The patent introduces an intermediary loss function that acts as a bridge between the neural network output and safety requirements. This loss function incorporates constraints (such as collision avoidance, dynamic feasibility, and legality) that guide the neural network to produce safe trajectories without requiring manual rule encoding. The intermediary loss function translates safety requirements into gradient signals that shape the network's learning process.
Solution Approach 2:
The patent changes the parameter space by formulating trajectory generation as an optimization problem with constrained parameters. Instead of directly outputting trajectory coordinates, the neural network optimizes parameters (weights, blending factors) that combine basis trajectories. This parameter transformation allows the system to maintain neural network flexibility while ensuring outputs satisfy safety constraints through parameter bounds and regularization terms in the loss function.
2Reliability
If rule-based methods are used for trajectory generation, then safety can be guaranteed, but the system requires significant manual effort and lacks adaptability
Solution Approach 1:
The patent replaces manual rule-based trajectory generation with a data-driven neural network system. Instead of manually encoding driving rules and heuristics, the system learns optimal trajectories from training data while safety constraints are embedded in the loss function. This substitution eliminates the need for extensive manual rule creation while maintaining safety guarantees through the constrained optimization framework.
Solution Approach 2:
The system enables self-service by allowing the neural network to automatically learn safe driving behaviors from data without continuous manual intervention. The loss function with embedded safety constraints guides the network to self-regulate its outputs, automatically adapting to different driving scenarios while maintaining safety requirements. This reduces the need for ongoing manual rule updates and adjustments.
3Measurement precision
If vast training data is collected and trained on, then neural network performance improves, but resource requirements become prohibitively high
Solution Approach 1:
The patent segments the trajectory generation problem into two parts: a neural network component that learns from data and a constrained optimization component that ensures safety. This segmentation allows the system to use moderate training data for learning general patterns while relying on the optimization framework with embedded safety constraints to guarantee correct behavior. The segmentation reduces dependency on massive training datasets by offloading safety-critical reasoning to the constraint-based optimization layer.
Data Source
AI summary
A method of generating at least one trajectory in scenario comprising an agent navigating a mapped area, the method comprising: receiving an observed state of the agent and map data of the mapped area; generating a set of multiple trajectory basis elements from the observed state of the agent based on the map data; processing one or more scenario inputs in a neural network to generate a set of weights, each weight corresponding to one of the trajectory basis elements; and generating a trajectory for the agent by weighting each trajectory basis element by its corresponding weights and combining the weighted trajectory basis elements.