Toward Autonomous Learning of an Ontology of Tool Affordances by a Robot

Jivko Sinapov, Alexander StoytchevAAAI 2008

Introduction

In robotics, the concept of "affordances" describes the relationship between an agent, an object, and the potential actions that can be performed. Traditionally, roboticists have pre-defined these outcomes (e.g., "the object moved left"). However, in complex, real-world environments, the number of possible outcomes is effectively infinite, and raw sensor data, such as high-dimensional video, is difficult to map directly to simple labels. This paper introduces a framework that allows a robot to autonomously build its own "dictionary" of actions and outcomes through incremental hierarchical clustering, discovering a taxonomy of outcomes and learning to predict them based on sensory input.

Methodology: The Affordance Ontology

The authors propose an Affordance Ontology designed to address outcome discovery and predictive modeling without human supervision. This ontology consists of two primary components:

  1. Outcome Taxonomy ($T$): A hierarchical tree structure where the root represents any change in the environment and the leaves represent specific, refined types of movement. The robot constructs this tree using an incremental version of the X-Means clustering algorithm. As a specific node accumulates enough data points, it "splits" into more granular sub-categories.
  2. Predictive Models ($\mathcal{M}$): At each non-leaf node, the robot trains an incremental ensemble classifier to predict which child node a trial will fall into. This prediction is based on the initial "cue" (a 30x30 retinal image of the object) and the chosen behavior.
The robot's experience cycle: (a-b) executing a behavior, (c) the visual cue, and (d) the resulting trajectory.

The robot's experience is formalized as $(B_i, C_i) \to O_i$, where $B$ is the behavior, $C$ is the visual cue, and $O$ is a 12-dimensional vector describing the trajectory of the object over time.

Experimental Setup

The study utilized a simulated environment created in the BREVE dynamic robot simulator. The setup included a simulated robotic arm equipped with a gripper and two distinct tools: a "T-Stick" and an "L-Stick." The target object was a small cylindrical puck.

The simulated BREVE environment featuring a robotic arm and tool.

The robot performed thousands of trials, generating its own internal dataset of visual cues and 12D outcome trajectories. This autonomous discovery approach eliminates the need for external "gold standard" benchmarks, allowing the robot to ground its knowledge in its own sensorimotor experience.

Key Insights and Results

Hierarchical Learning and Robustness

By structuring the prediction problem hierarchically, the robot avoids the complexity of distinguishing between dozens of outcomes simultaneously. It first predicts a general category (e.g., "moved left") and subsequently refines that prediction (e.g., "moved left and rotated"). In experiments with the L-Stick (1200 trials), the predictive models achieved an Area Under ROC Curve (AUC) between 0.82 and 1.0, with most models exceeding 0.9.

Autonomous Adaptation and Tool Transfer

The system is designed to be "open-ended." When introduced to a new tool (the T-Stick) that produces novel movements, the robot does not fail. Instead, it naturally splits its existing taxonomy to create new branches for the novel outcomes. For instance, it successfully identified a unique outcome—bringing the puck closer via rotation—and created a new leaf node ($v_{013}$) with a predictive AUC of 0.98.

Limitations and Future Work

While successful in simulation, the authors identify several areas for future development:

  • Physical Implementation: Moving from the BREVE simulator to a physical 7-DOF Barrett WAM robot arm to account for real-world sensor noise and physical unpredictability.
  • Environmental Complexity: Expanding the model to track multiple objects or more complex environmental changes beyond a single puck.
  • Behavioral Autonomy: While outcomes are learned autonomously, the six behaviors (push, pull, etc.) remain pre-defined. Future iterations could involve the robot learning the behaviors themselves.

Summary updated

BibTeX
@inproceedings{sinapov2008toward,
  author = {Jivko Sinapov and Alexander Stoytchev},
  title = {{Toward Autonomous Learning of an Ontology of Tool Affordances by a Robot}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure