Motion contains information about skill.
The best subject-independent model achieved a positive R², indicating that movement characteristics contain measurable predictive information.
103
Trials
3
Tasks
29
Engineered Features
6+
Regression Models
Research Question
Can characteristics of human movement be used to estimate skill?
Assistive robots may eventually need to understand how capable a user is before deciding how much assistance to provide.
This project investigates the skill-estimation component of that problem. Using kinematic data from the JIGSAWS dataset, movement characteristics were extracted and used to train machine learning models to estimate Global Rating Scale skill scores.
Methodology
Research Pipeline
01
Kinematic Data
Robotic movement trajectories from the JIGSAWS dataset.
→
02
Feature Extraction
Speed, acceleration, jerk, idle behavior, variability, and duration.
→
03
Machine Learning
Regression models trained to estimate continuous GRS skill scores.
→
04
Evaluation
Subject-independent validation on people not observed during training.
Dataset
103 trials across three robotic manipulation tasks
The JHU-ISI Gesture and Skill Assessment Working Set (JIGSAWS) contains robotic surgical kinematic data recorded using the da Vinci Surgical System.
Although the dataset originates in surgery, this project focuses on general characteristics of human motion rather than task-specific trajectory geometry.
Knot Tying
36 trials
Needle Passing
28 trials
Suturing
39 trials
Results
Best Model: Tuned Gradient Boosting
After evaluating several regression models and tuning the strongest nonlinear candidates, Gradient Boosting achieved the best overall subject-independent performance.
MAE
4.299
RMSE
5.193
R²
0.132
Actual versus predicted GRS skill scores under subject-independent evaluation.
Model Interpretation
What characteristics mattered?
Permutation importance suggests that movement variability, smoothness, and idle behavior contain useful information for estimating skill.
1
Left speed variability
2
Left jerk variability
3
Right idle ratio
4
Right median speed
5
Left acceleration variability
Key Findings
What the study showed
Generalization is difficult.
Performance on unseen individuals remains challenging, highlighting the importance of subject-independent evaluation.
Variability matters.
Speed and jerk variability were among the strongest individual predictors identified by the final model.
Simple features are not enough.
An R² of 0.132 shows that handcrafted summary features capture only part of the structure associated with human skill.
Future Work
From offline skill estimation to adaptive assistance
The next step is not simply to improve the regression score.
Future research can preserve the temporal structure of movement, evaluate skill across broader manipulation tasks, and ultimately investigate whether real-time skill estimates can improve adaptive shared-control systems.
Human Motion
→
Skill Estimator
→
Shared Autonomy
→
Adaptive Assistance
About the Research
NSF REU SITE: HUMANS MOVE
This project was completed during the Summer 2026 HUMANS MOVE Research Experiences for Undergraduates program at the University of Wyoming in the Robotics & Intelligent Systems Lab.
Researcher: Abdurrahman Oyediran
Computer Science · University of Southern Mississippi
Mentorship:
Dr. Chao Jiang
Umur Atan
Varun Bharadwaj