Behavior-Guided Path Planning for Vehicle-Agnostic Autonomous Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems for autonomous vehicle path planning require significant computational resources and manual labeling for each vehicle make and model, leading to inefficiencies and erratic control due to the need for separately trained deep neural networks (DNNs) and difficulty in converting 2D pixel coordinates to 3D world space.
Innovation Solution
A behavior-guided path planning system that uses automatically generated ground truth data to train a DNN, allowing it to predict paths through environments, eliminating the need for manual labeling and enabling vehicle-agnostic control by generating trajectory points in 2D or 3D world space without pre- or post-processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If end-to-end trained DNNs are used for path planning, then vehicle control can be achieved, but separate DNNs are required for each vehicle make and model, significantly increasing computational resources and training complexity
Solution Approach 1:
The patent applies universality by training a single DNN model that can generalize across multiple vehicle makes and models. The model learns vehicle-specific dynamics through behavior data collected from diverse vehicles during normal operation, enabling one model to perform path planning for heterogeneous vehicle fleets without requiring separate models for each vehicle type.
Solution Approach 2:
The system applies self-service by automatically collecting behavior data from vehicles during normal operation and using this data to train and update the DNN model without requiring manual intervention. The vehicles themselves generate the training data through their operational behavior, eliminating the need for separate manual data collection and model training processes for each vehicle.
2Measurement precision
If manual labeling or annotating is used to train DNNs, then training accuracy can be improved, but significant time and effort are required, especially as systems scale to different environments and vehicle types
Solution Approach 1:
The system applies self-service by automatically generating training data from vehicle behavior during normal operation. Sensors onboard vehicles collect data on steering angles, acceleration, deceleration, and environmental conditions, which is then used to train the DNN without requiring manual annotation. This automated data generation process eliminates time-consuming manual labeling while scaling efficiently to multiple vehicles and environments.
Solution Approach 2:
The patent introduces an intermediary process that automatically transforms raw sensor data from vehicle operations into structured training data. This intermediary system processes behavior data, extracts relevant features, and formats it for DNN training, replacing manual labeling operations and enabling automated model training across diverse vehicles and environments.
3Measurement precision
If annotations are made in image space, then correspondence to 2D pixel coordinates is achieved, but conversion to 3D world space is required, increasing computational costs and processing times
Solution Approach 1:
The patent inverts the conventional approach by not converting from 2D image coordinates to 3D world coordinates. Instead, it directly processes sensor data in a coordinate system that is natively compatible with vehicle control outputs, eliminating the computationally expensive coordinate transformation step while maintaining the necessary spatial relationships for path planning.
4Measurement precision
If DNNs are trained in image space, then 2D pixel coordinate paths are generated, but these paths may not correspond to desired 3D world space paths, resulting in noisy outputs and erratic vehicle control
Solution Approach 1:
The patent inverts the conventional training approach by training the DNN to directly predict vehicle control parameters (steering angle, acceleration, deceleration) from sensor inputs without intermediate 2D path generation. This eliminates the problematic coordinate space conversion and ensures that the model learns directly in the control-relevant parameter space, producing smooth and reliable vehicle control outputs.
Solution Approach 2:
The system applies parameter changes by transforming the DNN output from traditional 2D coordinate predictions to direct vehicle control parameter predictions. The model learns to output steering angles, acceleration, and deceleration values directly, changing the parameter space from spatial coordinates to control commands, which eliminates noise and improves control smoothness.
Data Source
AI summary
In various examples, a machine learning model—such as a deep neural network (DNN)—may be trained to use image data and/or other sensor data as inputs to generate two-dimensional or three-dimensional trajectory points in world space, a vehicle orientation, and/or a vehicle state. For example, sensor data that represents orientation, steering information, and/or speed of a vehicle may be collected and used to automatically generate a trajectory for use as ground truth data for training the DNN. Once deployed, the trajectory points, the vehicle orientation, and/or the vehicle state may be used by a control component (e.g., a vehicle controller) for controlling the vehicle through a physical environment. For example, the control component may use these outputs of the DNN to determine a control profile (e.g., steering, decelerating, and/or accelerating) specific to the vehicle for controlling the vehicle through the physical environment.