Apparatus for constructing map generation models and map generation apparatus utilizing the same

Robot-Captured Spatial Map Data Generation

This technology converts RGB and depth information from images captured by a mobile robot into embedding data via an encoder module. This data is then mapped with the robot's position information to construct grid-based spatial map data. Subsequently, a decoder module generates rendered images from this map, and by learning the differences from the original captured images through a loss function, optimizes the neural network-based map generation model.

Existing grid-based map generation methods suffer from decreased map accuracy due to the accumulation of robot localization errors. They also require significant memory for storing visual information and have slow data processing speeds, making them difficult to apply in real-world robot operating environments.

This technology introduces a deep neural network encoder-decoder architecture to embed features of captured images into a grid. Through efficient position-information-based data recording and rendering processes, it is an excellent technology that can improve real-time environmental perception and localization accuracy.

‍

‍

‍

Key Features:

• Applying captured images, taken by a mobile device moving through a learning space, to an encoder module to generate embedding data.
• Based on position information, recording embedding data into map base data to generate spatial map data.
• The map base data includes multiple grids where embedding data is recorded, and the embedding data includes RGB information and depth information for each pixel of the captured image.
• Comparing rendered images with captured images through a loss function to update the encoder and decoder modules and train the map generation model.

‍

This technology was developed with support from the Institute of Information & Communications Technology Planning & Evaluation (IITP) (SW Star Lab) research project 'Robot Learning: Efficient, Safe, and Socially Friendly Machine Learning'.

‍