On September 9 and 10, Jean-Claude Latombe attended the 1st quarterly TMR meeting in Cambridge, where he presented the current state of the project. The slides are available on the webpage of the project (http://underdog.Stanford.EDU/tmr/slides/) both in html format (click on Q1) and in ps format (click on 'postscript version). Important information communicated by Lt. Col. Blitch at this meeting was:

• Reconnaissance is the focus of the program. Clearly the 3 tasks that we consider (model building, target finding, and target tracking) are reconnaissance tasks.

• Dealing in real-time with moving obstacles (e.g., adapting paths to avoid collision) is critical for the program.

• Though indoor environments are the main types of environments considered in the TMR program, there is still some interest in outdoor environments.

• Reconfigurable communication network is an important topic to allow the reconnaissance robots to reliably communicate among them and with the operators.

We pursue our work to define realistic models/data structures for representing a the map of a building being scouted, for describing the perception and mobility capabilities in parametric form usable by our future software modules, for expressing the motion plans to be computed by these modules, and for defining a parametric model of a target's behavior. The key question that we consider is: What is the right level of complexity for the models? Trying to be realistic yields more complex models, but such models are likely to induce untractable planning algorithms. As models get more complex, planning algorithms will become more adhoc and heuristic in nature, with less clearcut performance guarantees.

We are working on extending the target-finding planner to use a more general visibility model than the idealistic line-of-sight one used in our previous planner. We are currently considering two extensions:

• In one extension, robot's vision is limited to a cone (e.g., a 45dg cone), instead of being omni-directional (that is, 360dg). A preliminary version of this extension has been implemented. Interestingly, the first experiments show that the new planner produces motion strategies that are more similar to those used by people (typically, the robot moves with its "back" close to walls) than the strategies generated by our previous planner with omni-directional vision (then the robot tends to follow the median axes of corridors and rooms).

• In the second extension, we explore the case where the robot would have downward vision. This case corresponds to the situation where the camera is placed at the top of a long pole mounted on the robot or where the robot is a helicopter (we could then imagine cooperation between ground robots and helicopters to find targets in a urban environment not limited to one building). No software is operational yet.

We also have started working on next-best-view techniques for model construction. No software is operational yet.

We have ordered part of the equipment that we will need to experiment with our software, including 3 mobile robots from Nomadic Technologies and PC and SGI workstations.

Dr. T.M. Murali has joined the team as a postdoc. Murali will become a major technical contact for Army liaisons and other groups in the TMR program.

Finally, we have started setting the webpage for our project. The webpage address is http://underdog.Stanford.EDU/tmr/.