Neural Actuation Mapping Using Visual Odometry for Adaptive Vehicle Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle control systems rely on static actuation maps that require interpolation and are not adaptable to changing vehicle dynamics, necessitating high-precision sensors and limited adaptability.
Innovation Solution
Implementing a dynamic actuation map using a neural network fed by visual odometry, which learns and updates actuation commands in real-time, eliminating the need for interpolation and high-precision sensors, and allowing adaptation to changing conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a static actuation map with look-up table is used, then the system structure is simple, but the adaptability to changing vehicle dynamics is poor and interpolation is required
Solution Approach 1:
The patent implements a dynamic actuation map using a neural network that continuously adapts to changing vehicle dynamics. The neural network model is trained offline with vehicle dynamics parameters and actuation commands, then deployed online to provide real-time adaptive actuation without requiring interpolation or recalibration when vehicle dynamics change.
Solution Approach 2:
The neural network is trained in advance during an offline phase using comprehensive training data that includes various vehicle dynamics conditions. This preliminary training enables the network to handle changing vehicle dynamics during online operation without requiring real-time recalibration or additional sensors.
2Measurement precision
If high-precision sensors such as GPS and inertial units are used, then the measurement precision is high, but the system complexity and cost increase
Solution Approach 1:
The patent replaces the need for high-precision mechanical sensors (GPS, inertial units) with a neural network-based system that computes dynamic set-points from camera images. The neural network processes visual information to determine vehicle dynamics and generates appropriate actuation commands, eliminating dependence on expensive precision sensors.
Solution Approach 2:
The patent introduces camera images as an intermediary medium to capture vehicle dynamics information. Instead of directly measuring dynamics with precision sensors, the system uses images as an intermediate representation that the neural network then processes to derive dynamic set-points and actuation commands.
3Manufacturing precision
If a look-up table with interpolation is used, then the implementation is simple, but the precision for target positions not in the map is reduced
Solution Approach 1:
The patent replaces the interpolation-based actuation map with a neural network that directly computes continuous actuation commands. The neural network outputs precise actuation values for any target position without requiring interpolation between discrete look-up table entries, maintaining both precision and implementation simplicity.
4Adaptability or versatility
If the actuation map is kept constant, then the system stability is high, but the adaptability to dynamics changes is lost
Solution Approach 1:
The patent implements a dynamic actuation map that adapts to changing vehicle dynamics while maintaining stability through proper neural network design. The network is trained offline to account for various dynamics conditions, providing stable yet adaptive performance during online operation without requiring frequent recalibration.
Data Source
AI summary
An apparatus includes an interface and a processor. The interface may be configured to receive pixel data corresponding to an area outside of a vehicle and sensor data from said vehicle. The processor may be configured to generate one or more outputs in response to the sensor data based upon dynamic set-points for the vehicle and one or more inferences made by executing a first trained neural network model. The processor may be configured to process the pixel data arranged as video frames. The trained neural network model may have been trained by processing the pixel data arranged as video frames, performing computer vision operations to detect features in the video frames, determining the dynamic set-points for the vehicle by applying visual odometry operations on the features detected in the video frames, and modifying a plurality of weights of an untrained neural network model based on the dynamic set-points, the sensor data, and a training dataset comprising the dynamic set-points and corresponding actuator values that are representative of one or more desired inferences.