Similarity-Aware Skill Reproduction based on Multi-Representational Learning from Demonstration

TL;DR

The paper tackles the problem of inconsistent skill generalization when relying on a single LfD representation by introducing a similarity-aware multi-representational framework (SAMLfD). It combines multiple representations (LTE, JA, DMP) with a variety of similarity metrics to construct region-based maps that indicate which representation best reproduces a skill under new boundary conditions, and it uses a classifier to infer performance in unseen regions. The approach is validated across three simulated and four real-world tasks, revealing both improvements in generalization and critical insights into metric biases, with implications for selecting representations and similarity measures according to task goals. Overall, SAMLfD enables more robust, interpretable, and user-guided skill generalization in robotics by explicitly comparing representation performance over the generalization space.

Abstract

Learning from Demonstration (LfD) algorithms enable humans to teach new skills to robots through demonstrations. The learned skills can be robustly reproduced from the identical or near boundary conditions (e.g., initial point). However, when generalizing a learned skill over boundary conditions with higher variance, the similarity of the reproductions changes from one boundary condition to another, and a single LfD representation cannot preserve a consistent similarity across a generalization region. We propose a novel similarity-aware framework including multiple LfD representations and a similarity metric that can improve skill generalization by finding reproductions with the highest similarity values for a given boundary condition. Given a demonstration of the skill, our framework constructs a similarity region around a point of interest (e.g., initial point) by evaluating individual LfD representations using the similarity metric. Any point within this volume corresponds to a representation that reproduces the skill with the greatest similarity. We validate our multi-representational framework in three simulated and four sets of real-world experiments using a physical 6-DOF robot. We also evaluate 11 different similarity metrics and categorize them according to their biases in 286 simulated experiments.

Figures (7)

• Figure 1: A flow diagram illustrating the proposed similarity-aware multi-representational LfD (SAMLfD) framework.
• Figure 2: All figures and tables presented in this paper follow this legend.
• Figure 3: Reproductions of letters $S$ and $\alpha$ using three LfD representations with different reproduction behaviors ranging from immediate convergence in JA to preserving shape in LTE.
• Figure 4: Generalization of three 2D skills using SAMLfD with three representations (See Section \ref{['subsec:simulated_exps']} for details).
• Figure 5: Generalization of four real-world skills using SAMLfD. For each skill, the experiment was repeated three times with a different input demonstration. The similarity regions can be seen above each test (see Section \ref{['subsec:real_world_exps']} for details).