Abstract
In human movement learning, it is most common to teach constituent elements of complex movements in isolation, before chaining them into complex movements. Segmentation and recognition of observed movement could thus proceed out of this existing knowledge, which is directly compatible with movement generation. In this paper, we address exactly this scenario. We assume that a library of movement primitives has already been taught, and we wish to identify elements of the library in a complex motor act, where the individual elements have been smoothed together, and, occasionally, there might be a movement segment that is not in our library yet. We employ a flexible machine learning representation of movement primitives based on learnable nonlinear attractor system. For the purpose of movement segmentation and recognition, it is possible to reformulate this representation as a controlled linear dynamical system. An Expectation-Maximization algorithm can be developed to estimate the open parameters of a movement primitive from the library, using as input an observed trajectory piece. If no matching primitive from the library can be found, a new primitive is created. This process allows a straightforward sequential segmentation of observed movement into known and new primitives, which are suitable for robot imitation learning. We illustrate our approach with synthetic examples and data collected from human movement.
Introduction
A central challenge in robotics is enabling machines to learn complex tasks by observing humans. While humans naturally chain simpler "movement primitives"—such as reaching, grasping, or stirring—into continuous sequences, robots struggle to parse these seamless transitions. This difficulty arises from two primary factors: co-articulation, where movements are smoothed together without distinct pauses, and the presence of unknown skills that are not yet in the robot's repertoire.
Current computer vision techniques offer statistical tools for segmentation, but they often lack the physical grounding required for motor control. This research bridges the gap by utilizing Dynamic Movement Primitives (DMPs) as a dual-purpose representation for both recognition and generation, allowing a robot to decompose human motion into actionable units.
Methodology: DMPs as Linear Dynamical Systems
The core of this framework is the reformulation of DMPs—traditionally weakly nonlinear differential equations—into a controlled linear dynamical system with Gaussian noise. This transformation allows the recognition task to be treated as a system identification problem solvable via probabilistic methods.
The segmentation pipeline operates sequentially:
- Candidate Point Identification: The system scans the trajectory for local minima in velocity and acceleration to find potential transition points.
- Matching via Expectation-Maximization (EM): For each segment, an EM algorithm (utilizing Kalman smoothing) evaluates the likelihood of existing library primitives. It optimizes hidden parameters, such as duration and spatial goals, to account for co-articulation.
- Unsupervised Skill Discovery: If no library match meets a likelihood threshold, the system assumes a new skill. It "completes" the trajectory segment and uses Bayesian linear regression to determine the optimal segmentation point, subsequently adding the new primitive to its library.
Experimental Evaluation
The approach was validated using both synthetic 2D sequences and real-world 3D human movement data.
Synthetic 2D Data
Using a digitizing tablet, the authors recorded sequences of basic shapes (semi-circles, lines, squiggles). The proposed stochastic DMP method correctly segmented 19 out of 20 sequences, significantly outperforming standard Dynamic Time Warping (DTW) baselines, which achieved only 50% accuracy.
3D Human Movement
A 20-second "coffee making" sequence (pouring milk, reaching for sugar, stirring) was captured using a 3D motion capture system. The algorithm successfully parsed these continuous actions. Notably, when a core primitive was removed from the library to simulate ignorance, the system identified the anomaly, learned the "new" skill on the fly, and correctly identified it in subsequent appearances.
Key Insights and Future Work
The primary contribution of this work is a dual-purpose representation that handles co-articulation naturally. By treating movement as a stochastic process, the system can infer the "true" intent behind smoothed human motions. Furthermore, the use of Bayesian model evidence acts as an "Occam's Razor," ensuring that newly discovered primitives are behaviorally meaningful rather than just mathematical artifacts.
However, the current greedy sequential approach is susceptible to cascading errors; an early misidentification can derail the entire parsing process. Future research will explore global optimization using Dynamic Programming to find the most likely sequence of primitives across the entire trajectory.
Summary updated