Robot Learning from Human Demonstrations: Handwritten Alphabet Trajectories and Human-Likeness Evaluation

How can robots learn to perform complex, multi-segment motor tasks—specifically handwriting—that are perceived by humans as natural and human-like? The authors aim to move beyond simple spatial accuracy to capture the nuanced dynamics of human motion, including timing and contact force.

The researchers collected a dataset of 3,142 handwriting demonstrations from 22 participants using a touchscreen teleoperation interface. They extended the traditional Gaussian Mixture Model and Gaussian Mixture Regression (GMM+GMR) framework to incorporate contact force and normalized time as additional dimensions. Furthermore, they adapted the model to handle non-continuous, multi-segment trajectories (e.g., pen lifts between strokes), which are common in handwriting but often cause issues for standard trajectory learning algorithms. The resulting motions were evaluated by 21 participants who rated the perceived human-likeness of the robot's output on a continuous scale.

The generated trajectories achieved an average human-likeness score of 71.50 (SD=22.56), suggesting that the framework successfully produces motion that is perceived as more human-like than robotic. Participants identified geometric positioning and the sequence of strokes as the most critical factors influencing their perception of naturalness. The study also provides an open-source dataset of handwriting dynamics, which serves as a benchmark for future research into human-robot interaction and imitation learning.

As robots move into shared human workspaces, their ability to move in predictable and natural ways is essential for building trust and facilitating collaboration. This work demonstrates that incorporating force and temporal dynamics into imitation learning models significantly improves the perceived quality of robot motion, providing a practical, interpretable alternative to more data-heavy deep learning approaches.