Expectation-Based Vision for Self-Localization on a Legged Robot

Daniel Stronger, Peter Stone

Introduction

Self-localization is a critical capability for autonomous robots, enabling them to navigate and interact with their environment. For legged robots, such as the quadruped Sony Aibo, this task is complicated by the "jagged" motion inherent in their walking gait, which introduces significant noise and instability into the camera feed. Traditional localization methods often rely on sparse landmark detection—identifying specific colored beacons or goals to triangulate position. This research explores an alternative "expectation-based" vision paradigm that utilizes the entire geometry of the environment to achieve superior localization accuracy.

Methodology

The study compares two primary approaches to self-localization:

1. Landmark-Based Object Detection (Baseline)

In this traditional approach, the robot identifies discrete landmarks (e.g., colored spheres or cylinders) within its field of view. By calculating the distance and bearing to these known points, the robot updates its pose estimate using a filtering mechanism like a Particle Filter or Kalman Filter.

2. Expectation-Based Vision (Proposed)

The proposed method leverages a 3D model of the environment to predict what the robot should see from its current estimated position. The process involves four key steps:

  • Prediction: The robot renders a wireframe of expected edges based on its last known pose.
  • Observation: Edge detection is performed on the live camera frame to identify physical boundaries.
  • Alignment: The system identifies correspondences between observed edges and the predicted wireframe.
  • Update: The robot's pose estimate is mathematically adjusted to minimize the distance between these features.
Initial misalignment between the expected 3D model (white lines) and the actual camera feed.
Successful alignment of the predicted wireframe to the observed environment after iterative updates.

Experimental Setup

The evaluation was conducted using a Sony Aibo ERS-7 robot on a standard 4.4 x 2.9 meter RoboCup soccer field. The robot's onboard CMOS camera (208 x 160 pixels) provided the visual input. To assess accuracy, the authors compared the robot's internal pose estimates against manual ground-truth measurements taken at 14 distinct locations across 10 experimental trials.

Results and Discussion

The expectation-based approach demonstrated a significant improvement in spatial accuracy over the landmark-based baseline.

Metric Object Detection (Baseline) Expectation-Based (Proposed) Improvement
Mean Position Error 11.81 cm 7.55 cm 56% ($p < 10^{-9}$)
Mean Angle Error ~7.2° ~6.5° Not Significant

The 56% reduction in position error highlights the advantage of using dense geometric data (lines and edges) over sparse landmarks. The iterative alignment process proved robust against the erratic camera motion of the legged platform. Notably, the entire expectation-based system was efficient enough to run in real-time on the Aibo's 576 MHz processor.

Conclusion

Expectation-based vision provides a powerful framework for fine-grained localization by treating the problem as a continuous alignment task rather than a discrete recognition task. While it requires a prior 3D model and a reasonable initial pose estimate, its superior accuracy makes it an ideal candidate for high-precision tasks. Future work may involve hybrid systems that combine landmark detection for global re-localization with expectation-based vision for precise tracking.

Summary updated

Abstract

This paper presents and empirically compares solutions to the problem of vision and self-localization on a legged robot. Specifically, given a series of visual images produced by a camera on-board the robot, how can the robot effectively use those images to determine its location over time? Legged robots, while generally more robust than wheeled robots to locomotion in various terrains (Wettergreen & Thorpe 1996), pose an additional challenge for vision, as the jagged motion caused by walking leads to unusually sharp motion in the camera image. This paper considers two main approaches to this vision and localization problem, which we refer to as the object detection approach and the expectationbased approach. In both cases, we assume that the robot has complete, a priori knowledge of the three-dimensional layout of its environment. These two approaches are described in the following section. They are implemented and compared on a popular legged robotic platform, the Sony Aibo ERS-7. Many researchers have addressed the problem of vision and localization on a legged robot, in part because of the four-legged league of the annual RoboCup robot soccer competitions (Stone, Balch, & Kraetszchmar 2001). Nevertheless, of all of the approaches to this problem on a legged robot, we are not aware of any that take the expectationbased vision approach. However, we show that this approach can yield a higher overall localization accuracy than a state-of-the-art implementation of the object detection approach. This paper’s contributions are an exposition of two competing approaches to vision and localization on a legged robot and an empirical comparison of the two methods.
BibTeX
@inproceedings{stronger2006expectation,
  author = {Daniel Stronger and Peter Stone},
  title = {{Expectation-Based Vision for Self-Localization on a Legged Robot}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure