Autonomous Control via Physical Simulation and Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional control methods for autonomous technical systems, such as robots, struggle to account for physical effects and consequences of their actions in the environment, leading to inefficient and unpredictable behavior.
Innovation Solution
A method combining physical simulation and machine learning to determine optimized action steps by simulating different combinations of actions using a physical simulation model that corresponds to the environment and initial state, allowing for the consideration of physical laws and effects like gravity and collisions, and training a machine learning routine to evaluate and select the most favorable action sequences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional control methods use path planning algorithms based on mathematical and numerical methods, then a feasible path can be determined, but the methods cannot accurately predict how actions will affect the environment due to physical laws
Solution Approach 1:
The patent creates a virtual copy of the physical environment through a simulation model that replicates physical laws and dynamics. This virtual environment allows the system to test and evaluate action sequences without affecting the real system, enabling accurate prediction of action consequences while maintaining adaptability to physical effects.
Solution Approach 2:
The patent performs preliminary simulation of action sequences in the virtual environment before executing them in the real system. By evaluating multiple candidate action sequences in advance through physical simulation, the system can select the most favorable sequence that accounts for physical laws, thereby improving prediction accuracy and adaptability.
2Adaptability or versatility
If physical simulation is used to simulate various combinations of action steps, then physical effects like gravity and collision can be considered, but the computation time increases
Solution Approach 1:
The patent divides the action planning process into discrete action steps that can be simulated and evaluated independently. By segmenting the overall task into smaller action sequences, the system can efficiently simulate and compare multiple candidates without requiring exhaustive simulation of all possible continuous trajectories, reducing computation time while maintaining physical accuracy.
Solution Approach 2:
The patent simulates a selected subset of action sequences rather than all possible actions. By using the machine learning model to prioritize promising action sequences and simulating only those, the system achieves adequate prediction accuracy with reduced computation time, avoiding excessive simulation of unlikely candidates.
3Productivity
If machine learning routines are trained on simulation data to determine optimized action sequences, then fast and precise path planning can be achieved, but the system requires extensive training data generation
Solution Approach 1:
The patent performs preliminary training of the machine learning model using simulation data generated in advance. By completing the training phase before deployment, the system transforms the time-consuming training process into a one-time setup activity, enabling fast real-time path planning during actual operation without repeated training delays.
Solution Approach 2:
The patent uses the virtual simulation environment to generate training data that mirrors real-world scenarios. This allows extensive training data to be generated efficiently in the virtual space without requiring equivalent real-world experimentation time, accelerating the training process while maintaining data quality.
Data Source
Figure 1
Figure 2
Figure 3A~3C
AI summary
The invention relates to a method and a device for the computer-aided determination of control parameters for a favorable action of a technical system, for example, a technical subsystem of a technical system. An initial state (AZ) and an environment (UM) of the technical system (TS) are detected, and a target state (ZZ) is input. A physical simulation model (SIM) of the technical system (TS), which at least partially corresponds to the environment (UM) and/or the initial state (AZ), is selected, and output values of the physical simulation model (SIM) are adapted to the environment (UM) and/or the initial state (AZ). Starting from the output values and with respect to the target state (ZZ), various combinations of action steps of the technical system (TS) are simulated, whereby control parameters of the technical system (TS) are varied to execute the action steps.Based on the simulation data, a machine learning routine is trained by evaluating each action step, and based on the trained machine learning routine, a combination of favorable action steps for the technical system (TS) is determined.