Robot Motion Control With Stable Jacobian Learning Across Tasks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for learning robot dynamics from demonstrations struggle to provide stability guarantees, especially in dynamical systems governed by neural networks, and lack flexibility in handling different control scenarios.

Innovation Solution

A method using a Conditional-NCDS approach that incorporates conditional variables into neural contractive dynamical systems (NCDS) to ensure stability and adaptability, employing a variational autoencoder (VAE) for low-dimensional latent space representation and a diffeomorphic decoder to maintain contractivity, allowing a single neural network to handle multiple tasks and skills.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If neural networks are used to learn robot dynamics from demonstrations, then the system can model highly dynamic motions efficiently, but stability guarantees become hard to provide

Engineering Contradiction:
Improveefficiency in modeling dynamic motionsVSAvoidstability guarantees
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes the parameter representation by using the logarithm of the Jacobian matrix instead of the Jacobian directly. This parameter transformation ensures that the learned dynamics inherently satisfy stability conditions while maintaining the flexibility of neural network modeling. The logarithmic parameterization converts the complex stability constraint into a simpler linear constraint in the parameter space.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

Instead of directly learning the Jacobian matrix and then ensuring stability, the patent inverts the approach by learning the logarithm of the Jacobian and deriving the actual Jacobian from it. This inversion allows stability to be guaranteed by construction through the matrix exponential function, which always produces positive definite matrices when given skew-symmetric inputs.

Inventive Principle:
Principle #13The other way round (Inversion)

2Measurement precision

If separate models are trained for different activities, then each activity can be optimized independently, but the system complexity increases

Engineering Contradiction:
Improveactivity-specific optimizationVSAvoidnumber of separate models
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates a universal neural network model that can handle multiple different activities and tasks through a single unified architecture. The model uses activity-specific parameter vectors that can be switched or combined to adapt to different tasks, eliminating the need for separate models while maintaining activity-optimized performance.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent introduces dynamic switching capability where the model can adapt its parameters based on the current activity. The parameter selection and combination are dynamically determined based on the input state and activity type, allowing a single model to flexibly serve multiple functions without requiring static separate models for each activity.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4686533A1Device and method for controlling a robot device
Publication Date: 2026.02.04 ROBERT BOSCH GMBH
  • EP4686533A1 patent drawingFigure 1
  • EP4686533A1 patent drawingFigure 2
  • EP4686533A1 patent drawingFigure 3~4

AI summary

According to various embodiments, a method for controlling a robot device is provided, comprising providing demonstrations for movements for each of a plurality of different activities, learning the Jacobian of a function representing the demonstrated dynamics by training a neural network to output, in response to input of information about a state of the robot device and one or more parameter values specifying the activity, a representation of a semi-definite matrix, which, when regularized to give a definite matrix approximates the Jacobian of the function representing the demonstrated dynamics of the robot device when performing the activity and controlling the robot device by, for a given state and a given activity to be performed, using the neural network to approximate the Jacobian of the function representing the demonstrated dynamics at the given state when performing the activity, determining dynamics according to the approximated Jacobian and following the determined dynamics.