Robot Localization by In-Place Rotation and Sequential Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous robots face challenges in accurately recognizing their position and orientation, especially during unexpected movements, which can lead to unsafe driving and failure in providing desired services due to the complexity of learning all possible paths and the reliance on deep learning models like PoseNet.
Innovation Solution
The implementation of a method where a robot can efficiently collect training data by rotating in place and using a trained artificial neural network to estimate its position or pose based on sequential images, simplifying the localization process and avoiding dangerous driving by not requiring identification of its initial position or orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the robot uses deep learning models like PoseNet for relocalization based on image features, then the localization accuracy can be improved, but the device complexity and computational requirements increase
Solution Approach 1:
The patent segments the localization task into two distinct phases: a learning phase where the robot collects training data by rotating in place to build a model of its environment, and an execution phase where the trained model rapidly determines position and orientation. This segmentation allows complex deep learning to occur only during training, while operational localization uses a simplified inference process, thereby reducing real-time device complexity while maintaining high accuracy.
2Reliability
If the robot learns all possible paths to accurately recognize position and orientation, then the localization reliability can be improved, but the productivity and efficiency decrease due to the complexity of learning all paths
Solution Approach 1:
The patent applies preliminary action by having the robot perform a comprehensive learning phase in advance, during which it rotates in place to collect training data from all possible orientations. This preliminary data collection creates a complete model of the environment that enables rapid, reliable localization during actual operation without requiring the robot to relearn paths during execution, thus maintaining high reliability while improving operational productivity.
3Speed
If the robot drives autonomously without accurate position identification, then the speed and responsiveness can be improved, but the safety and reliability worsen due to dangerous driving
Solution Approach 1:
The patent uses preliminary action by pre-training a localization model that enables the robot to rapidly determine its position and orientation with high accuracy. This pre-trained model allows the robot to maintain fast response speeds during autonomous operation while ensuring driving safety through reliable position identification, eliminating the need to choose between speed and safety.
Data Source
AI summary
A robot may include an input interface configured to obtain an image of a surrounding environment of the robot, and at least one processor configured to rotate the robot in place for localization, and estimate at least one of a position or a pose in a space, based on a plurality of sequential images obtained by the input interface during the rotation of the robot. The position or the pose of the robot may be estimated based on inputting the plurality of sequential images obtained during the rotation of the robot into a trained model based on an artificial neural network. In a 5G environment connected for the Internet of Things, embodiments of the present disclosure may execute an artificial intelligence algorithm and/or a machine learning algorithm.