Surface Recognition from Minimal Tactile Data

 Model-based recognition of an object typically involves matching dense 3D range data. The computational cost is directly affected by the amount of data for which a transformation needs to be found before carrying out the match against a model. We have been investigating on recognition of curved shapes using ``one-dimensional'' data generated by a touch sensor through tracking.

For recognition based on data registration, finding the rotation and translation of data points before superposing them onto a model typically involve a search in the 6D transformation space. We have reduced the degree of freedom to three by acquiring data points along three concurrent curves on an object. First, we locate their intersection point p on a surface model, which is determined by the values of the two surface parameters. At this location, we align the estimated object normal at p with the normal of the model and rotate the data curves about it through an angle to obtain their best superposition onto the model.

The quality of match is determined using a combination of table lookup and local optimization methods. First, the Gaussian and mean curvatures at a reference point p on the object's surface are estimated from tactile data. They are used in a table lookup to find multiple (discretized) candidate points on the model that have similar local geometries. Local searches are then performed starting at these points to register the tactile data onto the model. Recognition of the model depends on the quality of the registration in comparison with the results on other models.

Our method can recognize closed-form surfaces as well as free-form surfaces (which are represented as triangular meshes). For every surface model, a lookup table is constructed to store the principal curvatures pre-computed at points of discretization. Principal curvatures are obtained through differentiation for a closed-form surface and local parabolic fitting for a free-form surface. Registration of the data curves is performed via nonlinear optimization and by a discrete greedy algorithm for the two shape classes, respectively.

The figure below displays the result of registering three data curves (in black) onto the surface of an elliptic paraboloid. Their registered locations (in white) are very close to the original ones.



Closed-form Objects

In the our first experiment, we used the strategy to recognize a set of four algebraic shapes shown below. A total of 121 points along three concurrent curves were obtained on each shape and then matched against each model. For example, data from the first cylinder yielded matching errors (in millimeters) 0.033, 0.182, 0.230, and 0.271 against the four models in the left to right order. As a result the first cylinder was corrected recognized. So were the three other objects by such error comparisons.



Free-form Objects

We have also performed an experiment with ten free-form objects, four of which are displayed below, alongside their models. The big red dot on each object indicates the refrence point through which an Adept Cobra 600 manipulator sampled along three concurrent curves, and the other dots on the corresponding model are the candiate mesh vertices found after a table lookup. All data curves displayed in red color were successfully registered onto the models.



To recognize these free-form objects, we registered the three tactile data curves from each object onto ten models. A 10x10 table was constructed with rows and columns corresponding to models and objects, respectively. The (i,j)-th entry is the error of registering the data from the jth object onto the ith model. In every column, the diagonal entry has the smallest error, which implies that the correct model is recognized.

The results demonstrate that, even for model-based recognition of curved shapes, acquisition of dense range data is unnecessary. Though tactile shape sensing does not match the capability of global recognition on polyhedral objects, it has several advantages over 3D range sensing. First, it can identify the relative position and orientation of an object being manipulated by the robot hand. Second, range images are subject to occlusions of the camera or range sensor, which is not an issue for the touch sensor.

Initial work on recognition of algebraic shapes was presented at IROS 2006. A journal submission was recently made to include recognition of free-form shapes represented by triangular mesh models.


This research is supported by an NSF CAREER Award 0133681.

Last updated on Apr 30, 2008.

Yan-Bin Jia
jia@cs.iastate.edu