A Framework for Learning and Reusing Robotic Skills

Brendan Hertel,Nhu Tran,Meriem Elkoudi,Reza Azadeh

2024-06-28

Abstract:In this paper, we present our work in progress towards creating a library of motion primitives. This library facilitates easier and more intuitive learning and reusing of robotic skills. Users can teach robots complex skills through Learning from Demonstration, which is automatically segmented into primitives and stored in clusters of similar skills. We propose a novel multimodal segmentation method as well as a novel trajectory clustering method. Then, when needed for reuse, we transform primitives into new environments using trajectory editing. We present simulated results for our framework with demonstrations taken on real-world robots.

Robotics

What problem does this paper attempt to address?

The problem this paper attempts to address is: As robots gradually enter people's daily lives, users need an intuitive way to teach robots new skills. To achieve this goal, the paper proposes a framework for creating a library of motion primitives. This library can help users teach robots complex skills through Learning from Demonstration (LfD), and these skills can be automatically segmented into basic action units (motion primitives), which are then stored in clusters of similar skills. When these skills need to be reused, the primitives can be transformed into new environments through trajectory editing. Specifically, the paper mainly addresses the following three key issues: 1. **Discovery of motion primitives**: Automatically identifying motion primitives from complex demonstration actions. 2. **Clustering of motion primitives**: Clustering similar motion primitives together for easier management and reuse. 3. **Selection and reuse of motion primitives**: Selecting appropriate motion primitives in new environments and making suitable adjustments to fit new tasks. To solve these problems, the paper proposes the following innovative methods: - **Multimodal segmentation method**: Combining joint space, task space, and sensor data, using probabilistic models to segment demonstration actions. - **Elastic clustering algorithm**: A clustering method based on elastic maps that can automatically determine the number of clusters. - **Trajectory editing method**: Using Laplacian Trajectory Editing (LTE) to transform skills into new environments while maintaining the shape of the original trajectory. Through these methods, the paper aims to enable robots to learn and reuse new skills more efficiently and intuitively.

A Framework for Learning and Reusing Robotic Skills

A Robot Skill Learning Framework Based on Compliant Movement Primitives

Robot Complex Motion Learning Based on Unsupervised Trajectory Segmentation and Movement Primitives.

Dynamic Movement Primitives Based Robot Skills Learning

Learning Skills from Demonstrations: A Trend from Motion Primitives to Experience Abstraction

Moving Personal Robots in Real-Time Using Primitive Motions

Skill Learning Framework for Human-Robot Interaction and Manipulation Tasks

A Novel Skill Learning Framework for Redundant Manipulators Based on Multi-task Dynamic Movement Primitives

Learning Motion Primitives from Demonstration

CLFR-M: Continual Learning Framework for Robots Via Human Feedback and Dynamic Memory

Learned parametrized dynamic movement primitives with shared synergies for controlling robotic and musculoskeletal systems

A Human-Like Learning Framework of Robot Interaction Skills Based on Environmental Dynamics

A learning framework for semantic reach-to-grasp tasks integrating machine learning and optimization.

Human-Robot Skill Transfer with Enhanced Compliance via Dynamic Movement Primitives

Robot Learning Method for Human-like Arm Skills Based on the Hybrid Primitive Framework

Learning robot motor skills with mixed reality

Learning Reusable Manipulation Strategies

Robot Learning System Based on Adaptive Neural Control and Dynamic Movement Primitives.

SKID RAW: Skill Discovery from Raw Trajectories

Generalized Robot Learning Framework

Learning to Segment and Represent Motion Primitives from Driving Data for Motion Planning Applications