Introduction
For robots to effectively assist in domestic or industrial environments, they must be capable of acquiring new skills dynamically from human interaction. While Learning from Demonstration (LfD) has become a cornerstone of robotic skill acquisition, most current frameworks are "batch-oriented," requiring all training data to be available simultaneously. This poses a significant challenge for memory-constrained systems that need to learn incrementally over time.
Furthermore, many LfD methods struggle with full-pose trajectories, which encompass both 3D position and 3D orientation. Orientations, typically represented by quaternions, do not reside in Euclidean space; they exist on a curved Riemannian manifold. Standard linear operations like averaging cannot be applied directly to quaternions without violating their mathematical properties. This paper introduces a framework for incrementally building and refining a library of Movement Primitives (MPs) that correctly handles 6D pose data using Riemannian manifold theory.
Methodology
The proposed approach is built upon Via-Point Movement Primitives (VMPs). VMPs are advantageous because they allow a robot to generalize a learned motion while ensuring it passes through specific "via-points" to reach targets or avoid obstacles.
Riemannian Formulation and Incremental Updates
To handle orientations correctly, the authors treat the space of rotations as a hypersphere. Operations such as finding the mean or interpolating between poses are performed using geodesics—the shortest paths on curved surfaces. To maintain memory efficiency, the system stores only "sufficient statistics" (count, mean, and non-central moments) rather than raw trajectory data. This allows the robot to update its internal skill models using only the current model and the incoming demonstration, adhering to a strict definition of incremental learning.
Fundamental Operations
The framework defines seven core operations to manage the skill library across spatial and temporal dimensions:
- Adding/Improving: Initializing a new skill or refining an existing one with new data.
- Removing: Deleting skills that are redundant or no longer useful.
- Merging/Splitting (Modes): Combining similar skills or dividing a single skill into distinct variations (e.g., distinguishing between "reaching high" and "reaching low").
- Merging/Splitting (Temporally): Concatenating sequential actions or segmenting long actions into smaller components.
Evaluation and Key Results
The framework was evaluated using the KIT Bimanual Actions Dataset, which contains 6D hand pose trajectories of everyday tasks like pouring, cutting, and lifting. The experiments focused on the system's ability to detect via-points and adapt its library as new demonstrations arrived.
Mathematical Equivalence and Precision
A critical finding was that for operations like Adding, Improving, and Merging, the incremental update is mathematically equivalent to a batch update. The system achieved errors as low as $10^{-15}$ mm compared to batch processing, proving that memory efficiency does not come at the cost of accuracy.
Efficiency in Via-Point Detection
Finding consistent via-points across multiple demonstrations is computationally expensive. The authors introduced a Weighted Phase Average method that identifies these points in just 0.33 seconds, compared to 81.85 seconds for traditional brute-force searches, without sacrificing precision.
Adaptability through Mode Splitting
The system demonstrated high adaptability when encountering diverse demonstrations. When a single "retreat" skill was presented with significantly different trajectories, the framework successfully split the mode into two distinct primitives. This operation reduced reconstruction error by approximately 35-40%, allowing the robot to represent multi-modal behaviors more accurately.
Discussion and Future Work
While the paper establishes a robust mathematical foundation for incremental learning on manifolds, several challenges remain. The current implementation assumes "perfect segmentation," where the robot knows exactly when an action starts and ends. Future work will focus on integrating automatic temporal segmentation and developing high-level logic to help the robot decide when to trigger specific operations, such as splitting a mode versus simply improving an existing one. Additionally, while temporal merging and splitting were conceptualized, their full mathematical formalization remains a target for future research.
Summary updated