AFUN: Towards an Affordance Foundation Model for Functionality Understanding

Researchers introduced AFUN, an affordance foundation model for robot functionality understanding.

Given one RGB-D view and a task description, AFUN predicts a functional mask and a 3D post-contact motion curve.

It outperformed baselines on affordance segmentation, contact-point prediction, and motion prediction.

The model also worked on real robot manipulation without finetuning, showing strong open-world generalization.

A large standardized data pipeline across robot, human, simulation, and real-world data helped train and evaluate the system.