This week and this Friday Night Robotics I was working on a behaviour AI for the newly refurbished BubbleBoy! It is much easier to design an AI that you know will be interactable with thew orld! Without the headbobbing capability of BubbleBoy, this effort would not be worth it.
BubbleBoy’s behaviour is primarily focused on food and water. BubbleBoy lives its anthropomorphized life just be be fed/watered! This means that BubbleBoy will want to know when it will expect to be fed, so that way it can headbob at the most optimal times. Being fed/watered is having a button pressed on BubbleBoy’s green stage area (the part with the blue and white LCD).
BubbleBoy will have three sensors, and a rudimentary measure time. The three sensors are the LDR, lid switch, and Xbee. All of these are onboard the robot (except for the Xbee which isn’t implemented yet).
Here’s a broad flowchart of the behaviour, which I will then explain below in detail.
Creating the Expectation
This is the observatory phase. BubbleBoy initially does not have any expectation of when to be fed, so it waits around. While it is waiting, its collecting data from all of the sensors and storing them to an array. There has to be 10 numbers in the array for each sensor before BubbleBoy can proceed to the next stage, pattern finding. Once this is fulfilled, and if BubbleBoy is fed/watered, then it goes on to find a pattern.
BubbleBoy is seen as a simple robot. Thereby, its pattern finding is relatively simple as well. The main idea is to check each sensor’s array and see if there is a pattern within the residuals. Meaning, if looking for a sequential pattern, the array would be iterated through (starting at i=0, stopping after i=8), and i+1 would be subtracted from i. An average of the residual change would be calculated at the same time. The array would be iterated through again, this time to count how many items are within +- 10% of the residual average. If the count is above 7, then it is said that there is a sequential pattern in there.
The same process is done for a secondary pattern, and a ‘thirdary’ pattern. Meaning, every 2nd number and 3rd number is checked to see if there is a pattern. It also goes through and checks with offsets, just incase the pattern is “even/odd/whatever”.
If there are no patterns, a random primary sensor is chosen for BubbleBoy to work with.
A problem exists in determining which pattern for which sensor to trust the most. A thirdary pattern for a photosensor is less dependable than a sequential pattern for a lid switch. This is handled in the next step.
Determining the Sensor to Use
The sensor with whatever pattern to follow is chosen through a Bayes Filter. This allows for a simple specification of confidence levels for a given sense resulting in a particular state through a table where all rows add up to 1.0. The sense columns are the sensor and pattern type. Meaning, there’s three sensors for every sensor, giving 9 senses in total. The states are confidence intervals. Anything >= 85 would be considered the most confident.
The numbers in bold are the ones that are more probable to be a result. A sensor in particular to look at is the Xbee for a thirdary pattern. The numbers are dispersed in such a way that it is almost a bimodal distribution. In order to understand why this is, imagine the scenario in real:
Xbee modems can communicate for up to a few meters. If BubbleBoy only receives a message every 3rd instance, it could either be a clever pattern, or it could be a miscommunication.
For this reason, the largest probability is given to the least confident state, and the second largest probability is given to the most confident state. It depends on the Bayes Filter and random roulette for what will actually be chosen.
This is just for determining what sensor is the most confident, that will provide the best result. This can mean that the chosen sensor will be followed, but it also may not. This will be explained in the next step.
Calculating the Cost Adjustment
The adjusted cost is determined by using the result from the previous step, and multiplying the probability by the inverse of whatever column it was situated in. In simpler terms, if the result was found in the best confidence interval state, then it would be multiplied by 5. If it was in ( 85, 70 ], it would be multiplied by 4…
This is then used as the sense for the next Bayes Filter. The state is how much of a cost to reduce the sense by.
Once the cost adjustment amount is determined, it is then subtracted from the cost of the particular sense’s cell. All of the senses are in a grid, with the most “primary” being the closest to BubbleBoy. An A* search is then used to choose the closest and least cost sensor. Once this is done, BubbleBoy can get on to entertaining its audience while waiting for food/water!
Following the Expectation
This is the main loop of the program. It’s where BubbleBoy is thinking in the present time about if it will be fed or not!
The key idea here is that BubbleBoy is thinking in the now. Meaning, it tracks the patterns differently than it does when it is reflecting back on them (in a “past” thought).
Sensor data is retrieved and placed into an array of size 10. The data at i is checked wither it is within +- 10% of the average in the pattern. The threshold percentage amount differs for the type of pattern, where secondary would be +- 20%, and thirdary would be +- 30%. If the data fits in to the check, then a “yes” counter is incremented.
Once the “yes” counter is >= 6, the food level will begin to decrease by (i/2)^2. At the same time, BubbleBoy will begin to show signs that he is excited to be fed/watered soon, by spinning its hat and bobbing its head.
If the “no” counter is >= 6, then it means that the expectation isn’t really working in the present thought. A flag is set to redo the expectation once BubbleBoy receives food and water.
i is then either incremented or reset to 0, depending if it hit 9 or not. (That’s a sort of obvious step)
When the time elapsed from not receiving food exceeds 150% of that of the observed elapsed time, BubbleBoy goes in to a “wallow” mode. When BubbleBoy is wallowing, it spins its hat slowly, and bobs rarely.
If the time elapsed is ( 150%, 100% ], BubbleBoy is “angry” because it did not receive its food exactly before the time elapsed. The hat will not spin, and BubbleBoy will bob side to side, and once (quickly) in the opposite axis to simulate a sort of “twitching” to all this anger!
If the time elapsed is ( 100%, 85% ], BubbleBoy is eager to please. Hat tricks will be common, same as delightful bobbing. Depending on how much food/water BubbleBoy has, it may also hoola hoop!
Last part of Following the Expectation
Depending on when BubbleBoy was fed, if it was in ( whatever, 100 ], the expectation will be done. Essentially, BubbleBoy is a positive/eager thinker that believes it should always be fed before the elapsed observed time. What an attitude!
If it is in [ 85, 100 ), then the expectation will be kept.
To reformulate the expectation, the previous steps are executed on the collected data. Then, everything repeats!
What will be super interesting to see, in my opinion, will be when the discrepancies occur from past thought to present thought. It will be interesting to see which sensors fare better through that transformation.
It will also be interesting to see if this actually can work on an Arduino, and not in simulation. I created a simulated version of BubbleBoy in Processing earlier in the week.
I’ll do the initial coding and testing through this simulation, mainly because I already coded the Bayes Filter Algorithm (with random roulette) in Processing from 2009 Honors Summer Research. 😀 Plus, in Processing it is very simple to communicate to an Arduino through Firmata. I can read in the data from BubbleBoy through there.
Hopefully next Friday I will have devised a test sequence to test the soon-to-be-coded AI on. This will of course be Open Source, under the Attribution-NonCommercial-ShareAlike 3.0 Unported License (BY-NC-SA).
Let me know what you think of this AI in the comment section below! (And yes I know it is very linear, but BubbleBoy doesn’t have enough DOF in the real world to spend the effort making the AI more nonlinear, since the observed result will essentially be the same!)