What is it about?
Robots can learn new tasks by watching and imitating human demonstrations—a process called “learning from demonstration.” However, when people show robots how to perform tasks, their demonstrations often vary in smoothness, precision, and efficiency. This variability can confuse the robot and reduce its ability to learn. Our study defines measurable ways to assess the consistency of human demonstrations—how similar and stable they are across attempts. We developed ten simple motion-based metrics, such as path length, smoothness, and effort, that reveal whether a person performs a task in a reliable way. Using two robot experiments, we showed that more consistent demonstrations lead to significantly higher learning success and better generalization to new situations.
Featured Image
Photo by Snapmaker 3D Printer on Unsplash
Why is it important?
Most robot-learning research focuses on improving algorithms, but few studies ask whether the data provided by humans is good enough for the robot to learn from in the first place. This work fills that gap by offering the first systematic framework to evaluate demonstration quality before training begins. By checking the consistency of demonstrations, researchers and designers can filter or guide human input to ensure only high-quality data is used. This approach improves robot performance without changing the learning algorithm and paves the way for everyday users—not just experts—to effectively teach robots at home, in hospitals, and in workplaces.
Perspectives
Writing this paper was deeply rewarding because it connects human behaviour and robot learning in a tangible way. I was fascinated to see how subtle improvements in people’s demonstrations could dramatically enhance a robot’s ability to learn and adapt. I hope this work encourages others to think of robot learning not just as an algorithmic problem, but as a collaboration between humans and machines—where teaching quality matters just as much as model design.
Maram Sakr
University of British Columbia
Read the Original
This page is a summary of: Consistency Matters: Defining Demonstration Data Quality Metrics in Robot Learning from Demonstration, ACM Transactions on Human-Robot Interaction, December 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3773904.
You can read the full text:
Contributors
The following have contributed to this page
