Search papers, labs, and topics across Lattice.
This paper introduces A4D, a novel approach for robot planning that shifts from appearance-based reasoning to affordance-based reasoning, enabling robots to understand task-relevant functionalities of objects. By mapping visual observations into a functional latent space centered around affordances, A4D significantly enhances generalization to novel interactions, achieving a 94% accuracy on known affordances and over 90% on new affordances with minimal training data. The method also boasts a 100x increase in inference speed, marking a substantial improvement over existing state-of-the-art systems.
A4D redefines robot interaction by enabling planning based on what objects can do, not just how they look, achieving unprecedented accuracy and speed in novel scenarios.
Existing robot planning systems rely on appearance-based reasoning, where visual observations are encoded into latent spaces organized around object appearances (e.g., recognizing a "cart" based on how it looks). However, planning requires reasoning about task-relevant functionalities of objects (e.g., whether an object is "movable"), which appearance-based latent spaces do not capture. As a result, existing approaches struggle to generalize to novel robot-object interactions. We address this limited generalizability through affordance reasoning, enabling planning based on task-relevant object functionalities instead of appearance alone. We introduce A4D, which maps visual observations into a shared latent space structured around affordances (e.g., "movable"). By projecting visual observations into this functional latent space and measuring their proximity to affordances, A4D infers functionalities relevant to the observed object. Furthermore, we introduce an affordance discovery mechanism that expands the latent space to handle unseen scenarios where existing affordances are insufficient. A4D uses proximity in the functional latent space to quantify uncertainty in affordance inference and selectively triggers affordance discovery. We evaluate A4D across several planning tasks involving diverse and unseen affordances. A4D achieves 94% inference accuracy on existing affordances outperforming state-of-the-art approaches by over 15% points, improves new-affordance inference accuracy from 70% to over 90% with fewer than 10% of the original training data, and enables 100x faster inference. Code, videos, and data available at: https://A4Dance-reasoning.github.io.
