Robot Manipulation Skill Learning With Task-Parameterized Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robot systems face challenges in flexibility and re-usability due to hard-coded manipulation skills, which are not adaptable to new scenarios or varying conditions, and conventional training methods fail to account for environmental changes.
Innovation Solution
A method and system that utilize Task Parameterized Hidden semi-Markov Models (TP-HSMM) and Task Parameterized Gaussian Mixture Models (TP-GMM) to learn and reproduce manipulation skills from kinesthetic demonstrations, allowing for the identification of attached and free task parameters to enhance flexibility and generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manipulation skills are hard-coded for a particular task, then the robot system can perform the task reliably, but the system lacks flexibility and re-usability for new scenarios
Solution Approach 1:
The patent implements dynamics by transitioning from static hard-coded skills to dynamic learned skills. The robot system uses machine learning models (TP-HSMM and TP-GMM) that can adapt and update manipulation skills based on demonstrated trajectories, allowing the system to dynamically adjust to new scenarios while maintaining reliable execution through learned patterns
Solution Approach 2:
The patent applies parameter changes by transforming task parameters between different coordinate systems and adjusting model parameters during learning. The TP-HSMM and TP-GMM models learn to transform trajectories between world coordinate system and object coordinate systems, enabling the robot to generalize skills across different tasks and scenarios by changing parameter representations
2Ease of manufacture
If conventional demonstration training is used by simply recording and replaying trajectories, then the training process is simple, but the method fails when environmental conditions vary
Solution Approach 1:
The patent applies copying by recording demonstrated trajectories and creating learned representations (TP-HSMM and TP-GMM models) that capture the essence of the demonstration. Instead of simple replay, the system copies the skill pattern into probabilistic models that can generate varied trajectories suitable for different environmental conditions while maintaining the core skill intent
Solution Approach 2:
The patent implements dynamics by transforming static recorded trajectories into dynamic probabilistic models. The TP-GMM learns Gaussian distributions that represent variability in demonstrated trajectories, and the TP-HSMM models temporal dependencies, enabling the system to adapt to environmental changes while maintaining training simplicity through automated learning
3Adaptability or versatility
If task-parameterized models are used to learn skills from demonstrations, then the system achieves better generalization, but the model complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the skill learning into distinct modular components: TP-HSMM for temporal structure and state transitions, TP-GMM for spatial trajectory distributions, and separate handling of attached versus free task parameters. This modular architecture manages complexity by organizing the task-parameterized model into independent, trainable segments that can be combined
Data Source
AI summary
A method of training a robot system for manipulation of objects, the robot system being able to perform a set of skills, wherein each skill is learned as a skill model, the method comprising: receiving physical input from a human trainer, regarding the skill to be learned by the robot; determining for the skill model a set of task parameters including determining for each task parameter of the set of task parameters if a task parameter is an attached task parameter, which is related to an object being part of said kinesthetic demonstration or if a task parameter is a free task parameter, which is not related to a physical object; obtaining data for each task parameter of the set of task parameters from the set of kinesthetic demonstrations, and training the skill model with the set of task parameters and the data obtained for each task parameter.