Autonomous Driving Neural Network With MOE Task-Specific Heads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based autonomous driving solutions fail to consider the mutual impact and requirements of prediction, decision, and planning, leading to results that do not meet actual driving requirements, and lack generalization capability in diverse traffic environments.
Innovation Solution
A neural network architecture using a mixture of experts (MOE) network with parallel head networks for feature extraction and fusion, enabling independent optimization of each task while maintaining interaction, incorporating self-attention and cross-attention networks for enhanced feature extraction and fusion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a unified neural network is used for prediction, decision, and planning, then model generalization capability is improved, but task independence and optimization flexibility deteriorate
Solution Approach 1:
The patent divides the unified neural network into multiple independent head networks, where each head network corresponds to a specific autonomous driving task (prediction, decision, or planning). This segmentation allows each task to be optimized independently while sharing a common feature extraction backbone, thus maintaining both task independence and model generalization capability.
2Device complexity
If independent single-task encoding is used, then task independence is improved, but computing power consumption and latency increase
Solution Approach 1:
The patent merges the feature extraction processes of multiple tasks into a shared backbone network, eliminating redundant computations. The multiple head networks share the same input processing and feature extraction layers, significantly reducing computing power consumption and latency compared to completely independent single-task models.
3Device complexity
If existing neural network solutions are used, then model complexity is reduced, but the ability to handle mutual impact and requirements of prediction, decision, and planning deteriorates
Solution Approach 1:
The patent introduces dynamic interaction mechanisms between the head networks, allowing them to adaptively adjust their outputs based on the requirements and states of other tasks. This dynamic adjustment enables the system to handle the mutual impact and requirements of prediction, decision, and planning effectively, improving reliability without excessive complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A neural network, and an autonomous driving method and apparatus are provided. The neural network includes a mixture of experts MOE network and N parallel head networks. The MOE network is used to perform feature extraction and feature fusion on input data to obtain N types of feature data, where the input data includes ego vehicle data, obstacle data, road topology data, and navigation data, the N types of feature data are in a one-to-one correspondence with the N parallel head networks, and each of the N parallel head networks corresponds to one autonomous driving task (S310); and each head network performs one corresponding autonomous driving task based on one type of feature data that is correspondingly input, where the autonomous driving task includes a prediction task, a decision task, and a planning task (S320). Each autonomous driving task can maintain independence to some extent while maintaining certain interaction, which is more in line with an actual driving requirement.