Robot Gain Tuning for Precise Motion Under Force Variation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional approaches for controlling robots, such as deterministic algorithms and machine learning models trained using simulations, often fail to adapt to variations in real-world environments, leading to underperformance or robot failures, especially in high-precision tasks like inserting a peg into a hole with varying sizes and positions.

Innovation Solution

A computer-implemented method using two trained machine learning models: a force planner model generates robot motions and predicted forces, and a gain tuner model determines compliance gains to adapt to real-world conditions, allowing the robot to perform high-precision tasks robustly without requiring precise simulations of physical interactions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If deterministic algorithms with pre-programmed operations are used to control the robot, then the robot can execute tasks with high precision, but the system cannot adapt to variations in operating conditions such as different component sizes and positions

Engineering Contradiction:
Improvetask execution precisionVSAvoidadaptability to variations in operating conditions
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies dynamics by transitioning from static deterministic algorithms to dynamic machine learning models that can adapt their behavior based on real-time sensor feedback and varying operating conditions. The trained ML models enable the robot to dynamically adjust its control parameters when encountering different component sizes, positions, or unexpected changes in the environment, thereby resolving the contradiction between precision and adaptability.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent utilizes parameter changes by training machine learning models to automatically adjust control parameters based on input features such as component geometry, material properties, and environmental conditions. The models learn optimal parameter settings from simulation data and can generalize to unseen variations, allowing the system to maintain precision across diverse operating conditions without reprogramming.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If machine learning models are trained using real-world training data from physical robot interactions, then the models can learn to handle variations in the environment, but data collection by trial and error is time consuming and can be damaging to objects, robots, and sensors

Engineering Contradiction:
Improveability to handle variations in environmentVSAvoidtime for data collection
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent applies copying by creating virtual copies of the real-world environment through high-fidelity simulations. These simulations replicate physical interactions, materials, and objects, allowing the machine learning models to be trained in a virtual environment that closely mirrors reality. This approach eliminates the need for time-consuming and potentially damaging trial-and-error data collection in the physical world, while still enabling the models to learn robust handling of environmental variations.

Inventive Principle:
Principle #26Copying

3Loss of time

If simulations are used to train machine learning models, then data collection is faster and safer, but the trained model often cannot correctly control a physical robot when the real-world environment differs significantly from the simulation environment

Engineering Contradiction:
Improvetraining timeVSAvoidperformance in real-world environment
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The patent creates highly realistic virtual copies of the physical environment through detailed simulations that accurately model physics, materials, and object properties. By improving the fidelity of these virtual copies, the trained models can generalize better to the real world.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent implements feedback mechanisms where simulation parameters and environmental conditions are continuously adjusted based on discrepancies between simulated and real-world performance. This feedback loop allows the simulation environment to evolve and better represent reality, improving the transferability of trained models from simulation to physical deployment.

Inventive Principle:
Principle #23Feedback

4Manufacturing precision

If precise simulations of physical interactions are used to train machine learning models, then the models can control physical robots accurately, but the simulations are prohibitively computationally expensive and time consuming

Engineering Contradiction:
Improveaccuracy of robot controlVSAvoidcomputational expense
Core Design Contradiction:
Manufacturing precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by selectively simulating only the most critical physical interactions and environmental factors that significantly impact robot performance. Rather than attempting to model every detail of the physical environment with maximum precision, the simulation focuses on key parameters such as contact forces, friction, and geometry, achieving sufficient accuracy for effective control while reducing computational expense.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4528406A1Techniques for controlling robots using dynamic gain tuning
Publication Date: 2025.03.26 AUTODESK INC
  • EP4528406A1 patent drawingFigure 1
  • EP4528406A1 patent drawingFigure 2
  • EP4528406A1 patent drawingFigure 3

AI summary

One embodiment of a method for controlling a robot includes generating, via a first trained machine learning model, a robot motion and a predicted force associated with the robot motion, determining, via a second trained machine learning model, a gain associated with the predicted force, generating one or more robot commands based on the robot motion and the gain, and causing a robot to move based on the one or more robot commands.