Back to APL home page

Autonomous Exploration

This page is under construction! Please bear with us.

An Introduction

Passively accepting measurements of the world is not enough, as the data we obtain is always incomplete, and the inferences made from it uncertain to a degree which is often uncaceptable. If we are to build machines that operate autonomously they will always be faced with this dilemma, and can only be successful if they play a much more active role. Here we present such a machine.

The machine deliberately seeks out those parts of the world which maximize the fidelity of its internal representations, and keeps searching until those representations are acceptable. We call this paradigm autonomous exploration and the machine an autonomous explorer.

We have constructed a working autonomous explorer in our laboratory. The links below will take you through the exploration process step by step.

Data Acquisition

We have implemented the Autonomous Exploration paradigm on many different machines. For example, on a mobile robot: the robot is free to move around an object in approximately the same manner a human being would.

In our most popular implementation of the autonomous explorer, data is acquired through a laser range-finding system mounted on the end-effector of an inverted PUMA 560 robot manipulator. The system has a field of view of about one cubic metre which can be positioned anywhere in the robot workspace.

The laser range-finding device was codevelopped with the National Research Council.

The images below show sample range data rendered as depth map and a shaded range image.

Back to the top of this page

Visual Reconstruction and Data Fusion

Because of the relatively low positioning accuracy of the robot (on the order of 1.0 cm), the transformation parameters relating different viewpoints must be computed from the acquired data.

To facilitate estimation of these parameters and to provide the necessary structure from which to perform shape analysis, a visual reconstruction procedure is used to turn the discrete sampled data from the rangefinder into a piecewise-smooth (C2) representation of the surfaces in the scene. From there, data acquired from different vantage points can be fused by determining the correspondence between features in adjacent views.

Back to the top of this page

Shape Analysis and Parts Decomposition

At this level of abstraction, reconstructed surface information from multiple viewpoints is used to determine surface boundaries corresponding to the parts of an object. The perceptual basis of the algorithm is the Hoffman and Richards principle of transversality regularity.

Objects in the scene are represented as conjunctions of convex solid. Boundaries between parts of objects thus correspond to concave discontinuities and/or negative local minima in the principal curvatures of the surface.

The images below show how a side view of the owl gets segmented into different parts:

Another example of parts decomposition: the pencil sharpener range image was provided by the National Research Council.

Back to the top of this page

Volumetric Modelling

At the highest level of abstraction, surface regions defined by part boundaries are described by parametric models such as superquadrics. In addition to serving as a basis for the characterisation of uncertainty, these descriptions provide additional cues for maintaining correspondence at the level of parts, for describing general shape properties, and for recognition.

The images below show how volumetric models were fitted to each part found in the segmentation process, and how the level of confidence in the model varies over its surface. The lighter the gray, the more uncertain we are about the model's ability to reproduce reality.

Back to the top of this page

Gaze Planning

Coming soon...

Back to the top of this page

Last update to this page done 1996/1/18 by trep