Automated Sensor Data Labeling for Autonomous Robot Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training recognition models for object recognition in sensor data are manual, time-consuming, and prone to inconsistencies, limiting the volume and quality of training data available for autonomous robots.
Innovation Solution
An automated method for generating training data using a trained auxiliary recognition model that synchronizes and transfers object attributes from auxiliary sensor data to primary sensor data, allowing for the creation of high-quality, consistent training datasets across multiple sensor modalities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If manual methods are used to generate training data, then quality control can be maintained, but the process is time-consuming and limits the volume of training data
Solution Approach 1:
The patent uses a generated sensor data set (synthetic copy) instead of manually processing real sensor data. The generation model creates artificial sensor data with ground truth labels, which are then used to train the recognition model. This copying approach eliminates manual annotation time while providing unlimited training data volume.
Solution Approach 2:
The patent performs preliminary generation of training data using a generation model before the actual recognition model training. The generation model pre-creates labeled sensor data sets, which are then used for training. This preliminary action shifts the time-consuming task from manual annotation to automated synthetic data generation.
2Manufacturing precision
If manual annotation is used, then labeling consistency can be controlled, but the process is labor-intensive and produces limited data
Solution Approach 1:
The generation model performs self-service by automatically creating labeled sensor data without human intervention. The model generates sensor data and simultaneously creates accurate ground truth labels based on the simulated environment, eliminating the need for human annotators and ensuring consistent labeling throughout the training data set.
Solution Approach 2:
The patent creates synthetic copies of sensor data with known ground truth labels through the generation model. These copied data sets maintain labeling consistency because they are generated from controlled simulations rather than human annotation, while enabling unlimited data volume for training.
3Quantity of substance
If a single sensor modality is used, then the system remains simple, but the training data volume and diversity are limited
Solution Approach 1:
The generation model serves multiple functions: it generates sensor data for different sensor modalities (e.g., LiDAR, radar, camera), creates ground truth labels, and prepares training data sets. This multi-functionality allows the system to handle multiple sensor types without requiring separate processing pipelines for each modality.
Solution Approach 2:
The patent merges multiple sensor modalities into a unified training framework. The generation model combines data from different sensor types (LiDAR point clouds, radar signals, camera images) and their corresponding ground truth labels into integrated training data sets, enabling the recognition model to process multiple modalities simultaneously.
Data Source
AI summary
A method for generating training data for a recognition model for recognizing objects in sensor data of a sensor. Objects and object attributes are recognized in auxiliary sensor data of an auxiliary sensor mapping at least one overlapping area using a trained auxiliary recognition model, and the object attributes of the objects recognized in the overlapping area being transferred to the sensor data mapping at least the overlapping area in order to generate training data.