One-Shot Cross-Geometry Skill Transfer through Part Decomposition
Robots can now learn complex manipulation tasks from a single demonstration by breaking objects into parts.
A team from Brown University led by Skye Thompson, Ondrej Biza, and George Konidaris has published a breakthrough robotics paper titled 'One-Shot Cross-Geometry Skill Transfer through Part Decomposition.' The core innovation is a method that enables robots to learn manipulation skills from a single human demonstration and then apply that skill to objects with completely different shapes. This solves a major limitation where current robotic systems often fail when encountering objects with unfamiliar geometries.
The method works by first decomposing both the demonstration object and the novel target object into their constituent semantic parts (like handles, bases, or lids). It then uses data-efficient generative shape models to accurately map the key interaction points from the parts of the demonstrated object to the corresponding parts of the new object. The system autonomously constructs and optimizes an objective function to align these skill-relevant points, enabling successful transfer. The research, accepted for ICRA 2026, shows the method generalizes to a wider range of object geometries than prior work and works for various skills in both simulated and physical robot environments.
- Enables one-shot skill transfer: Robots learn from a single demonstration and apply it to novel objects.
- Uses part decomposition: Breaks objects into semantic components (e.g., handles, lids) for accurate point mapping.
- Outperforms existing methods: Achieves superior generalization across diverse object geometries in sim and real-world tests.
Why It Matters
Dramatically reduces the data and programming needed to teach robots new manipulation tasks in warehouses, homes, or factories.