This week we looked at the mechanical design for our Connect 4 robot. We need a way to hold the tokens and drop one into a slot when the game-playing AI decides what to play.
We started with two ideas for holding the tokens:

1. A vertical magazine with 21 tokens held in place by two pins.
2. A disc holding the tokens at the edge.

The photo below shows an illustration of the magazine idea.

The magazine would move across the board and drop a token into a slot by pulling out the lower pin. The upper pin would hold the rest of the tokens in place so that the wouldn’t drop into the board. Once the bottom token was dropped, the lower pin would be pushed back and the upper pin pulled out to allow the tokens to drop down one step.

The disc idea involved it rotating into place over the board and dropping a token into the board as needed.

We decided that both ideas were impractical on size grounds: the magazine and disc would be very high and hard to support.

We eventually settled on a modification of the magazine idea: 3 magazines with 7 tokens each. This would be smaller and more stable. It does require us to keep track of the number of tokens in each magazine. That’s something our code should be able to manage.

We decided to keep the position of the magazines fixed. A carrier (think cup) could move under the magazine, collect a token and drop it into the board. The magazine and carrier would both need a single pin at the bottom to hold the tokens in place. The carrier should be the height of a single token, so that only one can drop into it. The carrier would have an arm (like a handle) that sits on the magazine side and holds the other tokens in position.

We spent some time discussing how to move the pins in and out before one ninja had a great idea: the pin should be the arm of a servo motor. The servo can rotate the arm through 90° to allow a token to drop.

We then considered how to move the carrier back and forth. We settled on having a single stepper motor drive a gear on a rack that holds the carrier. Turning the gear moves the rack and the carrier into position. Keeping track of the carrier position might be difficult. The easiest way would be to measure the time it takes to reach a particular slot and then power the motor for that amount of time. The problem with that is that it depends on the motor always taking the same amount of time to reach a slot. The motor might slow down as the battery driving it runs down. Mentor Declan suggested using a beam sensor to locate the correct position. We could put a marker above each board position and stop the motor when the sensor reaches the marker. That marker could be something as simple as a small hole. If the beam shines through the hole instead of reflecting, the mark has been found.

Lastly, we considered how to hold everything in place. We think that a large piece of plywood should do the job. We can also attach a sheet of paper to the plywood behind the board to help the vision system see the empty slots. Mentor Kevin pointed out that anything in the background of the board with the same colour as the tokens could confuse the vision system.

The photo below shows a sketch of the final design.

So, where will the parts from our design come from?
We have an old flatbed scanner that we can scavenge parts from. It has a gear and rack system to move the scan head. It should be long enough to meet our needs. We haven’t looked at the motor that drives the gear yet. We will have to figure out if we can power it and control it. If not, we have motors from previous projects that we can use. The gear probably won’t fit one of the old motors, so we will have to modify it. Or, we can 3D-print a part that will allow us to fit the gear to the motor. That worked well for us last year when we made axles to allow us fit wheels to the motors we had for our hide and seek robots.
We will have to 3D-print the carrier. It will probably take a few iterations to get something that works. It will need to be the right size to hold only one token, and be shaped to stop tokens from dropping lightly and getting wedged. Some experimentation will be needed. We can 3D-print the magazines as well, or maybe make something from plywood strips and cardboard.
We will also need to make a camera mount for our vision system. The mount will need to sit in front of the board, out of the way of the human player. It will need to be in a fixed position so that it covers the whole board and allows us to reliably identify token positions. This will probably be a mix of a 3D-printed part and plywood.
We will tackle these tasks at our next session.

As we plan our AI robot to play Connect-4, we have decided that Python would be a good programming language for the job, as it is widely used for many of the tasks we will need to do:

• Computer vision
• Artificial intelligence
• Hardware control

Therefore, we spent time this week brushing up on Python. Following the same approach that we had used previously to move from Scratch to C, we looked at how we would move from programming in Scratch to Python.

Here is the full set of notes (PDF): CoderDojo-Hackers-IntroToPython

Kevin then spent time explaining how to write a basic Python program to read an image from a webcam. Here is the code:

# We are using OpenCV
import cv2

# Capture webcam image
camera=cv2.VideoCapture(0)
val, img = camera.read()

# Display the image
cv2.imshow("Window display", img)

# Stop using the camera
camera.release

The Goal

In this session, we started planning how we will build an intelligent robot that can play the game of Connect-4. Our aim is that it will be like playing against a human, with a physical Connect-4 set, not just a computer game. This is the plan we had come up with during our brainstorming the previous week, so this week we started to figure out how we can achieve this.

This will be a challenging project that the whole group will work on together. It will require lots of teamwork, collaboration, and learning.

The Major Components

We spent time thinking about the major components that our system will need, and planning the major tasks for these components on the whiteboard. They will include:

1. A robot mechanism to play a move, which will involve moving to one of the positions 1-7 at the top of the board and dropping in a token, then reloading for the next time.
2.  A camera to view the board and analyse what it sees, to figure out which spaces have a red token, which have a yellow token, which are free, and whether anyone has won (4 in a row) or has a promising state (one or more 3-in-a-row).
3. An AI decision-making system (see below)
4. A software version of the game that we will find online and modify so that our AI can play it, so that we can work on the AI before the physical robot is ready.

For each component, we identified some major tasks and people volunteered to work on 2 major components each.

We also agreed that Python is a good programming language for the task, so we will have to brush up on Python next.

The Artificial Intelligence

The group spent some time playing rounds of Connect-4 against each other, in order to get us thinking about how we would design a computer strategy to play the game. Then we returned to the whiteboard, where people gave their ideas about the main strategies to be followed – these are on the right side of the whiteboard.

Some of the strategy ideas people proposed:

• Work towards 4 in a row (obvious but important!)
• recognise states with the potential for 4 in a row
• have multiple paths to win
• try to block your opponent.

This made me think of a general AI algorithm for playing 2-person games, called the Minimax Algorithm. This is based on constructing a game tree of all possible future moves, by yourself and by your opponent, for a number of steps into the future (called the lookahead or depth limit). Then, we evaluate all possible future game states using a utility function (also called fitness function): this will return a low value (e.g. -20) if we lose, a high value (20) if we win, or some other intermediate score if the game is not over, such as a count of the number of 3-in-a-rows we have, minus the count of our opponent’s (this will definitely be between 20 and -20). We will then pick the move that should lead to us a state with as high a utility as possible. Also, if we assume our opponent is rational, they will pick the move that will give us as low a utility as possible, so we can use the game tree to predict what their best possible move is, and be prepared for them! Each time they take a move, we can re-calculate the game tree to plan a new best move.

We then looked at the strategy ideas people had proposed, and saw how they are achieved by the Minimax algorithm, even though it’s a general algorithm, not just for Connect-4!

There are some good videos and tutorials about the minimax algorithm on line. Here are two that I liked:

• https://towardsdatascience.com/how-a-chess-playing-computer-thinks-about-its-next-move-8f028bd0e7b1
• https://www.youtube.com/watch?v=l-hh51ncgDI

This week we modelled a soft-drinks can in SketchUp.  A typical can is 135mm high and has a diameter of 66mm.
We made the model in two halves with overlapping edges that can be pushed together.
This model could be 3D-printed and used to hold components for one of our electronics projects.  It gave us a feel for using SketchUp and designing our own models.

We drew a profile of one half of a can and extruded that profile around a circle to make the can. We included a lip on the inside of the shell in the profile. The detailed instructions to make the model are here.
One thing we discovered was that we need to delete the face in the circle, leaving just the outline, before extruding.  If you leave the face intact, the extruded model will be missing its bottom face.

Then we made a profile of the other half, with a lip on the outside.  We put the two together in SketchUp to see how they fitted.  Section planes are helpful for this: insert a  section in the vertical plane and move it in and out using the Move tool to see that the lips touch, but don’t overlap.

One ninja had a great idea:  SketchUp allows us to apply materials to faces.  So, he applied a glass texture to the model so that he could see through it and check that it was correct.

The method of drawing the profile in the instructions above is quite laborious.  The intention is to give us an understanding of SketchUp’s inferencing system for drawing in a particular plane and snapping to end-points and mid-points.  Our ninjas were able to construct the profile much quicker by drawing two touching rectangles, and then removing a section (to make the lip) with a third rectangle.

Next week, hopefully, we will try printing our models on the 3D printer to see how they look.  Our 3D printer software needs a file in STL format to print.  SketchUp doesn’t support STL out of the box. There is an extension in the 3D Warehouse that lets us export our model to STL, so we have to install that first.

At Hackers this week, we learned how to solder. Group members stripped wires and then soldered them together, and they made LED circuits by soldering them onto stripboard, and tested them with Arduino programs.

As we discussed, it is important to build your circuits temporarily with a breadboard (where you just push the wires in, and can easily move them) before moving onto soldering them on to stripboard. Stripboard (also called Veroboard) has holes every 2.5mm in a grid, and has copper strips on one side connecting the holes in one direction. You mount the component (such as an LED on the side with no copper, and solder its pins to the copper strip. Then, you can solder a pin of a different component somewhere else along the same copper strip, and current can flow through the copper strip.

There are plenty of videos on YouTube  to demonstrate soldering technique. Here is one by Emer Cahill of GMIT: