Robot Control Using Learned Object Motion Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot control systems struggle to predict and navigate around objects in dynamic environments without impairing their movement, especially in areas with incomplete or unknown maps, due to limitations in detection and simulation methods.
Innovation Solution
A machine learning system is trained using a map that includes a training action space to determine the movement patterns of objects, allowing it to learn and adapt to the environment by approximating movement variables, such as probability of occurrence, transition, and presence, which enables it to plan trajectories that avoid interfering with object movements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex simulations or calculations are used during robot operation to determine movement patterns of objects, then measurement precision is improved, but device complexity and computational resources increase significantly
Solution Approach 1:
The machine learning system is trained beforehand using simulated movement data to learn object movement patterns. During actual robot operation, the pre-trained system quickly infers movement patterns from sensor data without requiring complex real-time simulations, thus achieving high precision with reduced computational complexity
Solution Approach 2:
The patent creates a simplified internal representation (map with action space) that copies essential features of the environment. The machine learning system learns to predict movement patterns within this simplified model, avoiding the need to simulate complex physical realities while maintaining sufficient accuracy for navigation decisions
2Device complexity
If traditional calculation methods are used to determine movement patterns in unknown or incomplete map sections, then device complexity is reduced, but measurement precision deteriorates due to insufficient information
Solution Approach 1:
The patent introduces a machine learning system as an intermediary between sensor data and movement pattern determination. This intermediary has been pre-trained on simulated data and can infer movement patterns in unknown areas by recognizing patterns from the learned model, bridging the information gap without requiring complex calculations
Solution Approach 2:
The machine learning system changes the parameters it uses to determine movement patterns - instead of relying on complete environmental information, it uses a reduced set of key parameters (sensor readings, map sections, action space definitions) that it has learned are sufficient for accurate prediction, enabling operation in incomplete map sections
3Adaptability or versatility
If the robot adapts its movement to avoid interfering with objects, then social compatibility is improved, but productivity decreases due to additional control constraints
Solution Approach 1:
The machine learning system is pre-trained to recognize and predict object movement patterns. During operation, this pre-trained knowledge allows the robot to quickly adapt its trajectory to avoid objects without requiring complex real-time optimization calculations, maintaining both social compatibility and operational efficiency
Solution Approach 2:
The system continuously monitors object positions and movements, using the machine learning system to predict future object locations. This feedback loop allows the robot to plan trajectories that avoid objects while minimizing deviations from the optimal path, balancing social compatibility with productivity
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method (30) for operating a robot control system (13), which comprises a machine learning system (23), the machine learning system (23) being designed to determine, in accordance with a map (20) comprising an action space in which the robot (10) can move, a variable that characterizes a movement course of at least one object, which can move in the action space, in the action space of the map (20), said method comprising: providing at least one training map, which comprises a training action space in which the robot (10) can move; inputting a variable that characterizes the movement course of at least one object, which can be moved in the training action space, in the action space of the training map; training the machine learning system (23) in such a way that a variable determined by the machine learning system (23), which variable characterizes the movement course of the at least one object in a specifiable portion of the training map, approaches the input variable that characterizes the movement course of at least one object in the specifiable portion of the training map. The invention further relates to a computer program and to a device for carrying out the method (30) according to the invention and to a machine-readable memory element (12), in which the computer program is stored.