Mr. Berman called with the news that we didn't make it into the GC. He also told us we could work at the GC. All of our travel expenses and other expenses would be paid for.

I bought an 18 ah non-spillable rechargeable SLA battery from Batteries Plus for ca $60

I wrote up an explanation for the perception demonstration and sent it to Mr. Berman. This is what it said:

What you just saw is a demonstration of the perception system of our Grand Challenge entry, the Expeditor. This system uses neural network technology to process the image from a camera. The output of the networks can then be used to control actuators or motors; in this case they're controlling a servomotor with a paper arrow attatched. All I'm doing is pushing the cart in the direction that the arrow is pointing.

The Perception System is a very general tool that can be used for many purposes. This demonstration was put together in a couple hours and is only an example of one way it can be used. The core technology is a neural network software library that I wrote. The networks can be trained on any type of defined data; their inputs are not limited to visual information. Based on results from experiments such as this one, I am able to say that this system works - it automatically builds a neural network that produces the desired output based on training material provided to it.

I have written several software tools that help make the network-building process easier. The first of these tools, the "Information Collector" helps to create quality training information. First, it lets the user record information strait from the sensor - it is currently configured to record a video from the camera. During the demonstration, I recorded a video while guiding the cart down the hall and around the obstacles myself. The video itself is not enough to train the networks, however. The tool then allows the user to allocate portions of the video to specific responses that are desired of the neural network. During this demonstration, I selected sections of the video and marked them as "turn right" frames, "turn left" frames, or "go strait" frames. Finally, the video can be reduced. We reduced the video to 25x25 pixels.

The next software tool, the "network trainer", allows the user to set the specifications for a neural network and then builds the network using the information collected earlier. I used one neural network with thirty neurons in the input layer, one hidden layer with fifteen neurons in it, and an output layer with two outputs - turn left and turn right. Once the network has been configured, it trains on the information. It generaly cycles through the whole video about ten times and takes a total of about ten minutes to train a network.

The third software tool, the "robot builder", is where everything comes together. The trained networks are nothing but trained networks - they process the camera image and come up with results, but another program still has to decide what to do with the results. This tool lets the user write his/her own code to tell the computer what to do with the results presented by the networks, with built in support for controlling motors and other actuators. For this demonstration I wrote code that told the servo to point strait ahead if neither of the network's outputs were on, to point left if the first output was on, and to point right if the second output was on. This tool also lets the user program in GPS waypoints and all the information regarding that.

The final software tool, the "robot driver", has no user interface. When this program is run, the robot begins driving.

I put the perception system demonstration videos on Jeff's website at netventure.com and sent an email to Mr. Berman.

We had our site visit. We were able to show the inspectors the article, which mentioned them.