issues of noise and robustness. References to the traveling salesman problem were refreshing, but if the power, weight, and energy challenges for a single robot are significant, then it would seem that the task of choreographing a team of robots for which we lack remotely relevant battery or energy technologies is premature. These problems can become quite intractable with a moderate increase in problem dimensionality, and there is a need to explore super-computing platforms in this context, especially the parallel implementation of heuristic optimization algorithms to solve the larger scale problems. It is not clear whether a practical problem formulation is possible that would admit an exact solution as proposed by this preliminary work. Nevertheless, this has taken positive first steps to solving a difficult problem.
Machine Learning for Robotics
An overview presentation outlined the various ways in which machine learning is being used to improve robot performance. Using machine learning to tackle the challenging problems in robot intelligence is a good approach with many avenues for fruitful study. The advocated approach for control learning, which begins with a model for nominal control and then applies machine learning to optimize the control parameters, is an appropriate way of applying machine learning. Storing a database of experience to assist machine learning is a useful way to gather data for machine learning, although much theory needs to be developed to determine how to generate and identify the right data for a given problem. In the area of designing for learning ability, however, other than properly instrumenting the robot with sensors and monitors, the main ideas are not clear. The presentations lacked sufficiently concrete technical information on all of these learning topics. Because the civilian market is already doing a lot of work in this area, ARL should focus its efforts on the aspects of machine learning that are uniquely military in nature and not the focus of civilian efforts.
Developing Hybrid Maps to Promote Common Ground
This research attempts to develop topological maps in indoor environments that can be represented in human-understandable terms. To date, the main contribution has been the development of an online approach that works in cluttered environments. The approach involves analyzing point clouds from a Kinect sensor and decomposing the environment into regions and portals.
Generating human-understandable maps has potential benefit to Army applications. However, using the Kinect sensor just because it is available is not an adequate justification for the selected approach. This presentation could have been strengthened if a better justification for the selected approach had been given. Old techniques that look at the ceiling to determine the likely 2D footprint of the walls are most likely a better starting point. As with the other mapping research, this work would also likely benefit from increased conversations among the various performers who are researching localization and mapping.
The focus was on mapping Western structures and their contents—for example, opening doors, climbing stairs, identifying beds, tables, and chairs. In much of the third world, interior openings may be covered with a blanket or other hanging, ladders may be used to access other levels, and furniture may consist of cushions, mattresses, and the like. Even in modern buildings robots may be stymied by leaving the door open and hanging a blanket up in its place. Robots designed for Western architectures may also be stymied by hanging beads in windows and doors or by installing window screens. These problems should be addressed.