There are three main strategies to map building that his lab in Glennan is currently focused on. First, the robot is given an a priori floor plan and the scan is used to verify the location of the robot with respect to this floor plan. Second, there is no information beforehand, and the robot progressively builds an ‘absolute map’ (equivalent to our third person representation) consisting of the successive scans in the environment. Third, a mixture of the two such that the robot is given some coordinates or grouping of objects that will clue it in on where it is if it runs into it, akin to navigating by the stars. The sensors currently at our disposal are: lidar, far infared, color video, g.p.s., wheel encoder, yaw accelerator (sp?). The robots have the mobility of automated wheelchairs, and they are wi-fi enabled.

    Each different map building strategy has pros and cons. In the a priori scenario, some information that the lidar is picking up (because the lidar is the primary locative sensor) will be discarded because it doesn’t fit into the floor plan the robot is given. So, good information is going to waste. Furthermore, if all the information coming in does not fit to the floor plan, then essentially the robot will go ‘insane’ because the sensory data coming in doesn’t fit its preconceived notion of what it should look like so it rejects it and maintains it’s a priori information. This is naturalistic in the psychotic patients, but not necessarily regular humans. In this scenario there is a map and the robot is trying to place an ego marker on this map by comparing sensory data to encoded data.

    In the ‘robot in the wild’ scenario, the robot uses its lidar scanner to build up an absolute map and place an ego marker on this map. If gps is working (because sometimes under trees the signal is not too strong, apparently they absorb it), then it can get a coordinate estimate of its position and use that to help build the map. The same is true for the previous example except that usually scenario one is carried out in a building, which renders the gps unusable.

    In the combined scenario, let’s call it the ‘noble savage,’ the robot is given some initial structure, whether that be a certain way to interpret the data, for example vertical or horizontal lines (like navigating hallways) are important, or data constellations, for example this grouping of trees is here and once this specific grouping is recognized, then the robot knows where it is. He compared this to how people used to navigate by the stars, if the constellations appear in a certain part of the sky at a certain time, and then you know you must be at a certain location on the earth. This is a combination of building your own map and using some previous knowledge in order to locate yourself on that map.

    Overall, it seems that Newman’s mobile robots are working towards creating the third person representation, but working with that representation and using it to reason is where we come in. He was extremely encouraging in that respect, he basically said that his specialty was the mechanics and bottom up components, but he was hoping we could provide the higher order cognition. With this enthusiasm in mind, he also wanted to caution us to stay within the constraints that he put forth because no matter how brilliant our ideas, if they are not compatible with his hardware, then it doesn’t do either party any good.

    His robotic hardware is built for creating these ‘absolute maps’ in a variety of different approaches. I explained the scenario of ‘map transference’ between two agents that we have been discussing and he seemed very intrigued. He mentioned that the wi-fi capability of the agents would allow the ‘telepathic’ transfer of maps, which isn’t the most naturalistic, but in the beginning phases it would be fine.

    He also mentioned another interesting fact that the wi-fi allowed, namely controlling the robots from a remote location. We could be in the lab in 618 while collecting data from a robot in Dr.Newman’s lab on the first and second floor of Glennan. Or, if we prefer, we could run the software right on the robot from a mini-Mac unit that is installed on every robot.

    I mentioned the ideal versus real representation and tried to explain how this was at the heart of our approach. My reasoning was something like this: the agent creates a third person representation, additionally there is some input that causes a goal state representation, then the agent would try to simulate within a replica third person representation how to make the current map match this goal state map, creating an ideal action routine. Upon implementation, the action routine would change according to environmental feedback and evaluation of the current course of action with respect to the real representation. However, I think the explanation was somewhat problematic because I could tell that he was not exactly following. However, he did acknowledge the power of decoupling the ‘absolute map’ and performing mental simulation within this map, which he said would be possible after a few alterations to the software. So, essentially, we still need to show him exactly what reasoning with representations looks like, but it is possible based on the existing robotic structure.

    In his view, there are essentially two ways to get from a to b. First, go the way you know, and next, go the way you think is the shortest route. He gave an example of using the same roads to go home although they might not be the fastest, i.e. he knows he won’t get lost if he takes these ol’ roads. I think this issue is at the heart of the problem. Going the way you know is a first person, landmark based account of direction, while the shortest route is a third person, map based account of direction. The shortest route may require facing unknown territory. For example, you know where a and b are on a map, but a problem may come up somewhere in the middle. He used the example of a fence, or even the grand canyon, when is it more economical to stop following the fence in hopes of a gate before just turning back and starting over. How should we weigh one decision over another? This decision- making in action routines is an important avenue to explore. This will eventually lead us back to attention because one has to attend to something in order to decide on it.

    Dr. Newman is very interested in working with us, but he wants to make sure that our code is compatible with his robots. In order to ensure this, we have to take a few things into account. First, none of the existing robots have arms. He is kind of sad about this, and wants to let his robots literally ‘change the world’, but currently, this is not possible. Additionally, there aren’t touch sensors. Essentially, the robot won’t be doing a whole lot of interacting. However, it could run into things, and we could deduce the force from some math and physics base on how far it moved and acceleration of the robot, so we have an implicit force sensor.

    Essentially, the things we have to work with are mobility and perception. From this, we can create goal representations, attend, decide, categorize, mentally simulate, act, etc.