Robot Self-Recognition Using Conditional Probability-Based Contingency

Kevin M. Godby, Jesse A. LaneAAAI 2006

Abstract

As robots transition into dynamic social environments, the ability to distinguish self from non-self becomes critical. This paper introduces a robust statistical approach for robot self-recognition inspired by psychological theories of infant development. By implementing a "Conditional Probability-Based Contingency" method on a three-link robotic arm named Narcissus, researchers demonstrate how a robot can identify its own body parts by calculating the sufficiency and necessity of visual stimuli relative to its motor actions. The system proves resilient even against "anticipatory distractors"—external objects that mimic the robot's movements.

Introduction

A fundamental requirement for social robotics is a "theory of mind," starting with the basic ability to recognize one's own physical boundaries. Traditional methods often rely on simple temporal contiguity—checking if a visual event occurs immediately after a command. However, this is easily fooled by external objects moving in sync with the robot. This research addresses the challenge of self-recognition in the presence of such distractors by moving beyond simple correlation to a more sophisticated probabilistic model.

Methodology

The core of the proposed system is the Conditional Probability-Based Contingency method. This approach evaluates the relationship between a robot's behavior ($B$) and a perceived stimulus ($S$) using two primary indices:

  1. Sufficiency Index (SI): Defined as $P(S | B)$, it measures the probability that a stimulus occurs given that the robot performed an action. A high SI indicates that "when I move, this object almost always moves."
  2. Necessity Index (NI): Defined as $P(B | S)$, it measures the probability that an action was performed given that a stimulus was perceived. A high NI indicates that "when this object moves, it's almost always because I moved."
Visual markers on the robotic arm used for motion segmentation and contingency analysis.

The workflow involves an initial training phase where the robot performs random movements to establish the statistical distribution (mean $\mu$ and standard deviation $\sigma$) of time delays between commands and visual feedback. During detection, events falling within a $3\sigma$ window are correlated. An object is classified as "self" if both SI and NI remain above a threshold of 0.8.

Experimental Setup

The researchers utilized a custom hardware platform called Narcissus, a three-link planar robotic arm. The visual system employed a Firewire camera to track colored tennis balls attached to the arm's joints, facilitating motion segmentation.

Experimental setup featuring the Narcissus robotic arm and the visual tracking system.

The evaluation included standard runs to test basic recognition and a specialized "distractor" run. In the latter, a human experimenter moved a separate marker to intentionally mimic the robot's timing, testing the system's ability to differentiate between true self-motion and external mimicry.

Results and Discussion

The experimental results validated the effectiveness of the probabilistic approach:

  • High Accuracy: Narcissus correctly identified its own components with at least 85% accuracy across standard test runs.
  • Distractor Resistance: The Necessity Index proved to be the critical factor in defeating distractors. While a distractor might achieve a high SI by moving in sync with the robot, it inevitably moves at other times, leading to a low NI. In tests, the robot successfully labeled human-controlled markers as "other" while maintaining 100% "self" identification for its actual markers across 20 poses.

Conclusion and Future Work

This study successfully maps bio-inspired logic from human infant development onto a functional robotic algorithm. While the current system relies on simplified vision and 2D constraints, it provides a robust foundation for self-recognition. Future research will focus on expanding the system to 3D environments, handling self-occlusion, and improving visual processing to operate in complex, real-world settings without the need for high-contrast markers.

Summary updated

BibTeX
@inproceedings{godby2006robot,
  author = {Kevin M. Godby and Jesse A. Lane},
  title = {{Robot Self-Recognition Using Conditional Probability-Based Contingency}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure