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

Most robot Learning from Demonstration (LfD) work focuses on whether a robot completes a task correctly, not whether its motion looks natural to a human observer. The dominant algorithmic approach — Gaussian Mixture Models combined with Gaussian Mixture Regression (GMM+GMR) — encodes spatial trajectories from multiple human demonstrations into a smooth, generalized path. But existing implementations model only position (x, y), ignoring how hard a human presses and how they pace themselves through a stroke. Publicly available handwriting datasets like EMNIST or OnHW capture static images or unconstrained IMU recordings, neither of which is structured for robot trajectory learning. No prior work had paired a GMM+GMR system with a perceptual user study asking humans whether the output actually looks human-like — leaving a gap between algorithmic benchmarks and the HRI reality that motion style affects trust and collaboration.

The paper's core claim is that encoding contact force and normalized time alongside spatial position — and handling the pen-lift gaps in multi-stroke characters as separate, independently trained segments — is sufficient to produce handwriting trajectories that the majority of human observers perceive as closer to human motion than to robotic motion. In other words, the missing ingredient in prior GMM+GMR handwriting systems was not a more powerful model; it was richer input dimensions and explicit handling of trajectory discontinuities.

Twenty-two participants wrote all 52 uppercase and lowercase Latin characters on a 13-inch touchscreen using a capacitive stylus, guided by predefined templates and stroke-order prompts to keep demonstrations consistent. Two load cells sampled contact force at 80 Hz; a ROS system logged position, force, and timestamps. Each participant repeated every character three times, yielding 3,142 usable demonstrations after manual filtering. Each trajectory is represented as a four-dimensional state vector — x position (mm), y position (mm), contact force (N), and time normalized to [0,1] by average segment duration — making each training sample five-dimensional when the absolute timestamp is added. A GMM with K=20 Gaussian components (chosen empirically on held-out demonstrations) is fitted per stroke segment using the EM algorithm; GMR then retrieves a smooth consensus trajectory conditioned on time. Pen lifts are detected by a threshold on Euclidean distance in a weighted four-dimensional feature space; each stroke segment gets its own GMM+GMR, and segments are stitched in temporal order with linear transitions for the lift intervals. The result is one generated trajectory per character-case combination, resampled to a uniform 100 Hz grid. In the evaluation study, 21 different participants watched a simulated robot arm draw characters — marker thickness and transparency reflected the learned force profile — and rated each on a 0–100 continuous scale where 50 is the neutral midpoint between robotic and human-like.

Across 1,092 ratings, the generated trajectories scored a mean human-likeness of 71.50 (SD=22.56), and 81.2% of individual ratings exceeded the neutral midpoint of 50. Uppercase characters scored marginally higher (mean=72.0, SD=22.4) than lowercase (mean=71.0, SD=22.7). The highest-rated characters were 'o' (83.95), 'S' (82.38), and 'c' (81.76) — all geometrically smooth. The lowest were 'F' (57.43), 'j' (58.95), and 'k' (61.05) — characters with sharp corners, multiple strokes, or unusual stroke sequences. In the post-study questionnaire, participants rated geometric positioning (mean=3.71/5) and trajectory sequence (mean=3.57/5) as the most influential factors, while speed of motion was rated least influential (mean=2.81/5). The generated trajectories closely matched human demonstrations in path length (63.87 mm generated vs. 65.31 mm human) and mean force (1.084 N vs. 1.268 N), with smoother force profiles reflecting the regularizing effect of GMR.

This work establishes a concrete perceptual benchmark — 71.5/100 human-likeness — that future LfD methods for handwriting or planar manipulation can target using the released open-source dataset and generated trajectories. It also confirms that spatial accuracy and stroke sequencing matter more to human observers than speed, which should inform how researchers prioritize dimensions when designing motion quality metrics. What it does not answer: whether the same GMM+GMR approach scales to 3D manipulation, expressive writing styles like cursive, or real robot hardware, since the evaluation used simulation only with 21 participants in a controlled lab. The paper also does not compare against alternative LfD methods such as diffusion policy, so it is unclear how much of the 71.5 score comes from the GMM+GMR architecture versus the richer input dimensions alone. The key open question: would adding force and normalized time to a modern imitation learning baseline close the remaining gap to fully human-like perception?