Workshop: Steven McDonagh and Benjamin Rosman

Steven McDonagh

Simultaneous registration of many range images with robust density estimation

 Point datasets are routinely generated by optical and photometric range finders and have become increasingly popular in applications such as navigation, eHeritage and modeling from reality. Typical sensors can only measure the visible surface of the target, and therefore only provide a partial view of the object due to (self)-occlusions, blind areas or otherwise missing data. Generating high quality geometric representations from real-world objects requires the fusion of such partial views into a common coordinate frame by estimating the transforms between the datasets - multi-view registration.

Sequential approaches are not optimal because errors can accumulate and propagate. Moreover view sequences have to be known or manually specified in advance. We consider the problem of simultaneous global registration, where point correspondences and view order are unknown and the aim is to align all views simultaneously by distributing the registration errors evenly between overlapping viewpoints. Given many partial views, we estimate a kernel-based density function of the point data to determine an accurate approximation of the sampled surface. We use this density to guide an energy minimization in the transform space, aligning all partial views robustly. We evaluate this strategy quantitatively on synthetic and laser range sensor data where we find that we have competitive registration accuracy while improving convergence behaviour over existing frameworks for this task.

Benjamin Rosman

A Perception-based Action Abstraction for Multiple Tasks

Building robots capable of performing a number of different tasks in a complicated environment is a difficult problem. We are yet to see adaptable systems that can handle multiple tasks in rich worlds. This is largely due to the issues of incomplete information and partial observability. However, many real worlds have inherent structure which could be leveraged to limit the search for sensible behaviours.

To take a step in this direction, I present a model which allows a robot to represent and redescribe its environment in terms of its own capabilities. This involves integrating perception tightly into the control architecture, by extracting coarse topological scene descriptors which we leverage to simplify the incomplete information problem.  Exploiting this, the action-space can be described by a set of locally robust behaviours which are appropriate under the current perceptual signals. This representation naturally enables fast and reactive decision-making, flexible performance in multiple tasks, and replanning in response to unforeseen perturbations.

In this talk, I describe the progress already made in this direction with some simple experiments, and discuss an outline of how I will scale this to larger, more realistic settings.