Abstract
A fundamental challenge in developmental robotics is "grounding"—enabling a robot to learn meaningful categories from its own sensory experiences rather than pre-programmed labels. This paper presents a framework where a humanoid robot categorizes objects as containers or non-containers by integrating observations from six distinct exploratory behaviors. By employing consensus clustering to combine noisy data from individual interactions, the robot achieves a unified, highly accurate understanding of object properties.
Introduction
Traditional object classification in robotics often relies on static visual features or single-interaction types. However, biological agents use a diverse behavioral repertoire to explore novel items. A single action, such as grasping, may reveal size but fail to identify functional properties like containment. This research addresses how a robot can integrate multiple, potentially conflicting "opinions" from different exploratory behaviors—such as shaking, flipping, and moving—into a cohesive categorization.
Methodology
The study utilized an upper-torso humanoid robot equipped with two Barrett WAM arms and a 3D camera. The categorization pipeline consists of five stages:
- Exploratory Behaviors: The robot performs six actions: Drop Block (into/onto the object), Grasp, Move, Shake, Flip, and Drop Object.
- Feature Extraction: The robot tracks "visual co-movement" between the object and a small block, recording states where neither, one, or both items move.
- Visual Outcome Classes: Movement sequences are grouped into outcome classes using spectral clustering.
- Per-Behavior Categorization: X-means clustering creates initial object categories based on outcome frequencies for each behavior.
- Consensus Clustering: A hard consensus clustering algorithm finds a unified categorization that maximizes shared information across all six individual results, effectively filtering out noise from less informative behaviors.
Experimental Setup
The dataset comprised 20 objects: 10 household containers and children’s bucket toys, and the same 10 objects flipped upside down to serve as non-containers. The robot performed 100 trials per object for each of the six behaviors, totaling 12,000 interactions.
Results and Discussion
The results highlight that behavioral diversity is crucial for robust learning. Not all actions provided equal information; for instance, the "Flip" behavior was the most accurate, perfectly aligning with human labels, while "Grasp" was the most prone to error.
Individual behavior error rates were as follows:
- Flip: 0 errors
- Move: 3 errors
- Shake: 4 errors
- Drop Object: 5 errors
- Drop Block: 6 errors
- Grasp: 7 errors
Crucially, the consensus clustering algorithm successfully synthesized these results, achieving a final categorization with only 2 errors out of 20 objects. This demonstrates that the framework can derive meaningful, human-like categories even when the majority of exploratory behaviors are individually unreliable.
Conclusion and Future Work
This research demonstrates that "intelligence" in robotics is deeply tied to action. While the current model relies on visual co-movement, future iterations could incorporate auditory feedback (e.g., the sound of a block hitting a container) and active learning strategies to dynamically select the most informative next behavior.
Summary updated