Vax - Till I think of a better name

My entry into the club's (DARC) intelligent bot ball competition which requires a bot to find orange ping pong balls and return them to a goal.

My design places the guts of a Black and Decker Hand Vacuum, sometimes called a dustbuster, onto a mobile platform with a CMUcam for vision processing to find both the balls as well as the goal once it has filled up its tube with balls. The cam has a tilt servo so I can change the angle of view to see both down low where the balls are, and up higher later to find my goal.

The mobile base is a Sirius model kit from Budget Robotics (www.budgetrobotics.com). It is a tracked, differential drive platform with 2 servos driving the tracks.

Rev 2 added a servo controlled "gate" to the tube that allows the bot to "trap" balls once they're captured. That way the vacuum can be turned off and the balls still be held. This saves on battery time. I also added an IR photoreflective sensor so that the bot knows when it has a full tube.

This project still under construction, but a first pass at the design and photos seemed in order.
A before shot. This is the 4.8V model of the hand vac made by black and decker. Essentially I sawed off everything but the blue part, and I also sawed off the handle of that. This side view shows the blue vacuum part which contains a DC motor and the air impeller assembly. I kept the blue assembly since the impeller rotates inside it and all the mechanicals for mounting the impeller and the motor and the place where the suction does its business were all there, so i cut away anything i didn't need.

 
This rear view shows where I cut away the handle. you can see the terminals for the motor where i have wired yellow and green wires that feed to the hand soldered relay control board which is bolted to the bottom of the vacuum assembly. A DARC board is mounted below this and provides the brains. You can see the DC motor inside the housing and the hole in the top where the on/off thumb button was. This close up shows where I attached a 3 inch plastic tube cut with a bevel to get the angle I needed. Some careful usage of duct tape provides a very effective seal with little loss of suction.
Close up of the end of the tube. This shows the servo in "gate closed" position holding a ball. The IR photoreflective sensor board is bolted to the inside of the tube to get the sensor as close to the ball as possible. The sensor's output is about 4.95 volts when no ball is present and about 3.3V when a ball is present. Black electrical tape on other side from sensor helps to get a good strong voltage for no ball present. The gate is just some paper clip wire. View of the circuit boards. The relay board is the brown board bolted to bottom of vac assembly. The green DARC board below it controls via a single digital I/O line. The relay board just turns on/off a relay to feed battery voltage to the motor. I only rotate the motor one way, so I only need one relay. I found a transistor switch like a TIP120 caused some current loss and reduced performance of the vac suction. So I used a relay instead.
Theory of Operation: The DARC board sends a serial string to the camera telling it to track the color of the ping pong balls. The camera then looks for pixels in a frame of video (26 fps) that match that color range. For the orange balls in good light this is an RGB value (Red/Green/Blue) of 240/150/20 or so. You can specify a range for each so I pad it with +/- 20. The camera then responds with a serial string that gives an X/Y coordinate for where that color's center mass occurs in the frame. The software then drives the bot forward, or turns to close with the ball. Once it is close enough, the vac is switched on and stays on till the tube is full. Once that happens, the camera is tilted up so that the bot can see the goal, navigate to it, line up with the goal hole and the turn off the vacuum. Gravity does the rest.
Here is a frame capture from what the camera sees when I have it tilted to look down the tube. This frame capture is taken from the Java GUI which shows where in the frame the ball occurs. This frame capture shows the ball closer, in range for suction. Notice the center of mass has moved down the frame and to the center, and further the bounding box of pixels that match is also larger.

I divide the frame into 9 zones. Three rows (horizontally) and three columns (vertical). In the first photo, the ball is in zone 1 (top left corner), meaning that the bot has to move forward and to the left to get the ball near the suction tube. In the next frame the ball is in zone 7, close to 8. The software will maneuver the bot to get the ball into zone 8 of the video.

Still a lot of work to do but initial testing is sound. The suction is very good, enough to grab a ball that is 1 inch away from the tube, probably even more if I didn't mind waiting longer. This video shows the bot in scanning mode and picking up balls. Once it has 5 it closes gate, turns off vac, waits a few seconds then opens the gate. Click here to see a quick video of the suction action.