RoboFlag was invented in January of 2001 as a resource not only for Cornell but as a testbed that could be used to study numerous algorithms by Universities and Corporations the world over. It was formally introduced to the controls community at the 2003 American Controls Conference where it is very well received in the conference's first Interactive Session. Cornell has continued to be at the forefront of developing new and innovative technologies thru the testbed as well as continuing to expand the testbed itself.

Unique to the RoboFlag testbed, the system does not require an overall "brain" to manage the vehicle agents' interactions. Instead every agent is able to run its own separate program and artificial intelligence. Cooperative efforts are therefore invoked through inter-agent communication and negotiation allowing for a decentralized approach to creating the teaming behaviors. In addition, the RoboFlag testbed also incorporates an interface for multiple human operators to act as managers.

The overall goal of the program at Cornell is to develop generic widely applicable algorithms that will allow 'M' human operators to effectively manage 'N' agents in a highly dynamic environment where N >> M. The hierarchical system that supports this typically includes path planning at the lower levels, reactive decision making and teaming at the middle level, and resource management at the highest level. At every level, the system also allows a human operator the ability to intervene at decision points which has also led Cornell to compare and contrast various levels of autonomy and human control in its studies as well as investigate methods for improving human situational awareness.

Recent efforts have particularly focused on the teaming level. In the past, within the research community the focus has been more on the development of one particular robot that could be optimized and programmed with a special AI to accomplish a certain task. It was soon realized however that this approach had several downfalls. One of the largest was that if that robot was asked to accomplish a different task the robot either had to be significantly reconfigured or would perform at a significantly sub optimal level. In addition should that robot breakdown, the entire system it was a part of would be brought to a halt until the robot could be repaired. These problems have lead us to focus on the new approach of decentralized control.

Now instead of having one robotic agent, we replace that one with several independent smaller ones of varying capabilities with an artificial intelligence to communicate and discuss amongst themselves which agents should handle which part(s) of the task. Not only does this result in a more modular and flexible system but should one agent have to be removed from the overall system, the other agents can adapt and take on some of the removed agent's role.

This is not merely a conjectured plan however. We have already devised and tested several methods that have allowed us achieve behaviors of this level. In fact several papers have been written and proposed to the 2004 American Controls Conference detailing our efforts and the latest findings are scheduled for wide release in mid-2004 version of the RoboFlag Testbed.

To find out the details of some of our research please visit our Publications page.

To see the robots in action and to download conference posters please visit our Media Archieve.

To learn more about the inner workings of the test bed:

• The rules for the standard game release can be downloaded here.
• Details on the testbed's released software can be found in Documentation.
• An overview of our real robot Vision System can be found here.
• An overview of our real robot Wireless system can be found here.
• To download the latest version of our software release please visit Download TestBed.