Robot Arm Torque Control Using Learned Latent Dynamics

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing robot arm control methods fail to accurately compute joint torques, especially when nonlinearities such as friction and actuator forces impact movement, leading to imprecise control and performance issues.

Innovation Solution

A method using trajectory generators to create latent representation vectors from position, velocity, and acceleration data, which are then multiplied by trained core tensors to compute joint torques, incorporating basis functions trained on data sets to capture inherent robot arm properties, allowing for both feedforward and feedback control.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If rigid-body formulation is used to compute joint torques, then computational efficiency is improved, but control precision deteriorates due to inability to account for nonlinearities

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidcontrol precision
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent transforms the control approach by changing from fixed rigid-body parameters to adaptive learned parameters. Neural networks are trained offline to capture nonlinear dynamics, enabling the system to adapt torque computations based on actual robot behavior rather than idealized rigid-body assumptions.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces traditional mechanical control models (rigid-body dynamics) with a data-driven neural network model. This substitution allows the system to capture complex nonlinear behaviors that cannot be represented by conventional mechanical formulations, improving precision while maintaining computational efficiency through pre-trained models.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Manufacturing precision

If nonlinear regression techniques are used to account for friction and actuator forces, then control precision is improved, but computational complexity increases

Engineering Contradiction:
Improvecontrol precisionVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by performing complex neural network training offline before actual robot operation. During runtime, the pre-trained network provides fast torque predictions without requiring real-time complex computations, thus achieving high precision control while maintaining low online computational complexity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the control process into two distinct phases: offline training phase where complex nonlinear relationships are learned from data, and online execution phase where pre-trained models provide fast predictions. This segmentation transfers computational burden from runtime to setup time, resolving the contradiction between precision and real-time complexity.

Inventive Principle:
Principle #1Segmentation

3Manufacturing precision

If high feedback gains are used to compensate for control inaccuracies, then tracking precision is improved, but energy consumption increases

Engineering Contradiction:
Improvetracking precisionVSAvoidenergy consumption
Core Design Contradiction:
Manufacturing precisionVSUse of energy by moving object

Solution Approach 1:

The patent uses preliminary action by pre-training neural networks to accurately predict required torques before execution. This advance preparation enables precise feedforward control that reduces tracking errors proactively, minimizing the need for high-gain feedback corrections and thereby reducing energy consumption during operation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent converts the potential harm of inaccurate torque prediction into benefit by using neural networks to learn from past errors and improve predictions. The system transforms previous control inaccuracies into training data that enhances future performance, reducing both tracking errors and the energy required for correction.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Data Source

PatentEP3424650B1A method and apparatus for performing control of a movement of a robot arm
Publication Date: 2022.12.21 SIEMENS AG
  • EP3424650B1 patent drawingFigure 1~2
  • EP3424650B1 patent drawingFigure 3
  • EP3424650B1 patent drawing

AI summary

A method for computing joint torques applied by actuators (2) to perform a control of a movement of a robot arm having several degrees of freedom, DoF, the method comprising the steps of providing (S1) by a trajectory generator trajectory vectors, x, specifying a desired trajectory of the robot arm for each degree of freedom, DoF; mapping (S2) the trajectory vectors, x, to corresponding latent representation vectors, a, which capture inherent properties of the robot arm using basis functions, A, with trained parameters; multiplying (S3) the latent representation vectors, a, with trained core tensors, G, to compute the joint torques, t, for each degree of freedom, DoF.