Mobile Robot Traversability Detection Training with Simulated Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing mobile robot systems face challenges in accurately identifying traversable and untraversable regions in an environment without requiring large amounts of manually labeled training data, as manual labeling is tedious and prone to errors, and existing methods for automated labeling are noisy due to odometry miscalculations and external events.
Innovation Solution
A method is developed to train a machine learning model for traversability detection using synthetic images generated from a virtual environment, where a virtual robot is simulated to create high-quality training data with precise label masks, leveraging 3D models, sensor simulations, and collision checking to label traversable and untraversable areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manually labeled training data is used, then training accuracy is improved, but time consumption and cost increase
Solution Approach 1:
The patent creates a virtual copy of the real environment using 3D models and simulation software. This virtual environment replicates real-world conditions including lighting, textures, and physical properties, allowing the system to generate unlimited synthetic training images without manual labeling while maintaining training accuracy through realistic simulation data
Solution Approach 2:
The system performs preliminary actions by pre-defining traversable and untraversable regions in the virtual environment before generating images. Label masks are created in advance through simulation setup rather than post-processing, enabling automatic generation of labeled training data without time-consuming manual annotation
2Productivity
If automated labeling based on bumper sensors is used, then manual labeling is reduced, but labeling accuracy deteriorates due to odometry errors
Solution Approach 1:
The patent replaces the mechanical bumper sensor system with a virtual collision detection system in simulation. The virtual robot's collision detection uses precise digital models of objects and robot geometry, eliminating mechanical sensor noise and odometry errors while maintaining high labeling efficiency through automatic simulation execution
Solution Approach 2:
The simulation environment acts as an intermediary between the real world and the labeling process. It mediates by creating a controlled virtual space where collision events can be detected with high precision through digital modeling, avoiding the errors inherent in direct physical sensing while still enabling automatic labeling
3Length of stationary object
If LIDAR sensors are used, then detection distance is improved, but ability to distinguish hazard types deteriorates
Solution Approach 1:
The patent merges multiple sensing modalities within the simulation environment - combining LIDAR-like distance detection with camera-like visual information and physical property data. This integrated approach allows the system to detect objects at distance while simultaneously identifying their material properties, textures, and physical characteristics for accurate hazard classification
Solution Approach 2:
The system applies local quality by providing different types of information for different regions of the environment. Traversable regions are labeled with surface properties suitable for robot movement, while untraversable regions are labeled with hazard-specific properties, enabling the model to learn local distinctions between different hazard types based on their unique visual and physical characteristics
Data Source
AI summary
A system and method are disclosed for training a machine learning model configured to determine a traversability of a real-world environment by a mobile robot based on an image of the real-world environment. The method advantageously generates high quality synthetic training data for training a traversability detection model to discriminate between traversable and untraversable regions in images captured of a real-world environment. The synthetic training data is generated through simulation of a virtual robot in a virtual environment. Once the traversability detection model is trained, it can be deployed to the mobile robot for the purpose of predicting traversability of a real-world environment.