The machine just follows rules.
Just looking at the video, I guess the rules are these:
Follow a pattern (like a drum machine) for the tempo and rhythm of the notes.
Follow a pattern based on a setting as to whether the arpeggio goes up and repeats, goes down and repeats, or goes up and down.
Change the arpeggio only at the beginning of each measure.
At the beginning of the measure:
4a. if only one key is pressed, arpeggiate a major chord with that note as the root.
4b. If two notes a major third apart are played, arpeggiate a major chord with the lower note as the root.
4c. If two notes a minor third apart are played, arpeggiate a minor chord with the lower note as the root.
4d. If more than two notes are played, arpeggiate that chord.
Again, I'm just guessing, there can be more complex rules. The system could allow you to set the key, and then build a chord only on notes within the key. In this way, in the key of C, playing an A would result in an A minor chord.
An even more complex system could have a database of examples and use a Markov-chain or Bayesian method. These are a strictly mathematical technique, and (surprisingly) can be much easier and more effective than programming musical theory into the machine. Markov chains were used in the 1950's to program computers to automatically compose music.
Those methods work something like this: Suppose the key is C, the last time the chord changed was three measures ago, the current chord is G7, and the user presses E. In the past, under those circumstances, if the new chord had an E, 90% of the time it was C major, 8% it was A minor, 2% it was C# minor. For predictability, pick C major every time; for automatic composition, pick C major with 90% probability, A minor with 8% or C# minor with 2%.
Most likely the Roland arpeggiator does NOT do that, but if you like to program and have a system that has a MIDI keyboard for input and can play MIDI output you can try writing programs to do this.
Postscript: You may find it surprising that pure math can work as well as or better than music theory for doing something like this, but this just shows that there was a good reason that in ancient and Medieval times Music was considered to be a branch of Mathematics.