Robotic intelligence is defined as whether the robot knows its location, what path to choose, and what the surrounding threats and opportunities are. This key grouping of capabilities needs to be addressed by technologies that have to be developed for autonomous robotic operation. The presentation of this work showed slow but incremental progress in this area. Because this area is very challenging, ARL has appropriately been applying many approaches to finding solutions. However, many of these approaches are not fully coordinated with other approaches in the same area, and therefore researchers have not leveraged the lessons learned and emerging results of the other efforts. For example, it was mentioned that robots and warfighters do not actually collaborate, which makes the role of the warfighter in this project unclear. Nevertheless, some of the approaches are sufficiently developed for standardization—for example, mapping inside a building should now be standardized, and additional research should be aimed at object identification inside the building. ARL should lead a mapping effort to ensure that all areas are covered by the research and that maximum leverage is being obtained from the different approaches.
Advancements and Accomplishments The robotics enterprise is addressing some critical sub-problems whose solutions are necessary for increasing robot intelligence (e.g., mapping, path planning, machine learning, robot control, and architectures for cognition). The team has made several incremental advances, such as modeling the multi-robot patrol problem in a new way and using machine learning in a variety of ways to improve robot intelligence. The quadrotor control is very impressive; however, it relies on some strong simplifying assumptions (e.g., having perfect localization from a fully instrumented laboratory) that would make it very difficult to apply to real Army missions. It is difficult to point to particular advances that could change the game in terms of robot intelligence for Army applications, but this research is still solid, incremental work.
Opportunities and Challenges ARL should better define robot intelligence as it relates to warfighter needs. For example, what type of decision-making capability is needed? Relatively little was communicated during the presentation about the types of intelligence and decision making that are required to achieve the vignettes that have been outlined. A better vision of what is needed for Army applications would help to focus the research in intelligence. In particular, ARL should identify the forms of robot intelligence that are uniquely required for military applications but are not addressed by the significant amount of civilian work being done on robot intelligence.
For most of the research in intelligence, it is difficult to measure the accomplishments and progress. It is often not clear how the research fits into the state of the art or how the problems being researched directly affect an Army mission objective. More work can be done to benchmark the research, including the definition of short-, mid-, and long-term benchmarks of Army relevance. More solid demonstrations that are compared to benchmarks would support the case that relevant progress is being made. Also, the autonomous robot challenge is an excellent opportunity to demonstrate how to empirically measure progress. For example, when successful autonomous operation range increases from a few hundred yards to a few miles, measurable progress has been made.
The researchers should better elucidate the fundamental limitations of the models being developed, as well as better address issues of uncertainty and robustness throughout all the research tasks. This comment has broad applications. For example, in the MAST CTA, several of the intelligence projects and integrated demonstration did not acknowledge the size, weight, and power constraints or the sensing