Pooling Unit for Variable Road User Motion Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence units for automated vehicles face a trade-off between computational workload and expressiveness due to fixed-size inputs, which can lead to imprecision in representing environments, especially with low-resolution grids being inefficient and high-resolution grids being redundant.
Innovation Solution
An artificial intelligence unit, such as a reinforcement learning or Q-learning unit, is developed with a pooling unit that transforms variable input states of other road users into a single output state, allowing for permutation invariant transformation and adaptation of knowledge configurations to select optimal motion actions for automated vehicles, using a transformer unit and neural networks to evaluate and execute motion actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If fixed-size inputs are used in artificial intelligence units, then computational workload is reduced, but expressiveness and precision in representing the environment deteriorates
Solution Approach 1:
The patent applies dynamics by transforming the fixed-size input architecture into a variable-size input architecture. The pooling unit dynamically adjusts the number of input states based on the actual number of road users detected, allowing the AI unit to process varying amounts of environmental data without being constrained by a fixed grid size. This resolves the contradiction by making the system adaptable to different computational needs while maintaining precise environment representation.
Solution Approach 2:
The patent changes the parameter of input size from fixed to variable. By allowing the number of input states to vary according to the number of road users, the system can optimize computational workload dynamically while maintaining high expressiveness. The pooling unit processes a variable number of input states and produces a fixed-size output, enabling the system to handle different scene complexities efficiently.
2Measurement precision
If high-resolution grids are used, then environment representation precision is improved, but computational workload increases due to redundancy
Solution Approach 1:
The patent extracts only the relevant information from the environment by using a variable number of input states corresponding to actual road users. Instead of processing a complete high-resolution grid with all its cells, the system identifies and processes only the states that contain meaningful data about road users. This eliminates redundant computations while preserving the precision needed for accurate environment representation.
Solution Approach 2:
The patent applies partial action by processing only the necessary portion of environmental data. The pooling unit processes a variable number of input states that are sufficient to represent the environment accurately without processing unnecessary redundant information. This partial processing approach maintains precision while reducing computational workload compared to processing complete high-resolution grids.
3Productivity
If low-resolution grids are used, then computational workload is reduced, but expressiveness in representing the environment deteriorates
Solution Approach 1:
The patent makes the input architecture dynamic by allowing the number of input states to vary. This enables the system to adapt its processing capacity to the complexity of the environment, maintaining high expressiveness when needed while reducing computational workload when the environment is simpler. The pooling unit dynamically adjusts processing based on the number of road users, resolving the contradiction between efficiency and expressiveness.
4Device complexity
If a fixed grid size is used, then device complexity is reduced, but adaptability to different numbers of road users deteriorates
Solution Approach 1:
The patent creates a universal input architecture that can handle any number of road users through the pooling unit. The pooling unit serves multiple functions: it processes variable numbers of input states, maintains permutation invariance, and produces a fixed-size output suitable for the AI unit. This multi-functional design provides adaptability to different road user numbers while maintaining relatively simple device complexity.
Solution Approach 2:
The patent introduces dynamics into the input architecture by allowing the number of input states to vary while maintaining a fixed output size. The pooling unit dynamically processes different numbers of road user states and transforms them into a consistent format for the AI unit. This dynamic approach provides adaptability without significantly increasing device complexity.
Data Source
Figure 1~2
Figure 3~4
AI summary
An aspect of the invention describes Artificial intelligence unit for an automated vehicle, with said artificial intelligence unit comprising a pooling unit, configured to transform input states of a variable number of other road users in the environment of the automated vehicle into one output state, a knowledge configuration, said artificial intelligence unit configured to determine an evaluation value for at least two motion actions for the automated vehicle considering the output state and considering the knowledge configuration, select one of the motion actions considering the evaluation value of the respective motion actions.