Robots Learn Manipulation Tasks With Just a Few Demonstrations

Researchers have developed a method that lets robots intelligently ask for human demonstrations one at a time, then assess when they have enough examples to confidently execute manipulation tasks. Rather than requiring large datasets of pre-recorded demonstrations, this approach helps robots self-determine their learning needs.

How It Works#

The system combines three key components:

• Screw geometric representation: A mathematical framework that translates human demonstrations into executable manipulation plans, making it possible to measure whether a collection of examples is sufficient
• PAC-learning optimization: A sampling strategy borrowed from multi-armed bandit research that evaluates the robot's capability to generate plans across different areas of its workspace
• Weakness-targeting heuristic: A mechanism to identify gaps in the robot's learned skills and request new demonstrations from those specific areas

Testing the Approach#

A team conducted a user study with 22 participants who had no robotics background. Volunteers taught robots two practical tasks: pouring and scooping. The results demonstrated that fewer than 10 demonstrations were typically sufficient for the robot to learn and plan each task reliably.

Why This Matters#

The traditional programming-by-demonstration approach has long struggled with two fundamental questions: how do you know when you have enough examples, and how do you systematically find the ones you're missing? This work addresses both, potentially making robot training more efficient and accessible to non-experts.