I have been enjoying salsa (and other Latin dancing) for a number of years, and one of the things that has fascinated me is the subtlety and connection in human interaction that dancers have.

I have wondered if this subtlety is something that could be taught to robots?

The interactions in dancing are a microcosm of the interactions that humans have and show a complexity that I think is difficult for AI to learn. Looking at the example of dance can show us problems when applying robotic-type thinking to what may be considered inappropriate areas.

Who Is In Charge?

Salsa is typically a paired dance, with a man and a woman dancing together. In time with the music, the man will indicate to the woman which moves are to be performed by way of applying small amounts of pressure, and his dance partner will follow.

For example, if the man raises his left hand (while holding his partner’s right) and pushes forward on it, he indicates to her that she is to spin clockwise. However, if he raises his left hand and pulls, he indicates to the his partner that she is to spin counter-clockwise.

Through various types of leads the he can indicate all sorts of moves ranging from simple to extremely complex.

Because the man is leading, should we therefore say that he is in charge of the dancing? 

I don’t think so.  

Because he isn’t just telling his dance partner what to do in accordance with what he thinks she should do. He also needs to pick dance moves that are suited to his partner. For example, if she is a beginner, he should not try and attempt complex moves that she cannot accomplish. If she is a more advanced Salsa dancer, he will need to lead moves that are entertaining for her ability. He will also need to consider picking moves that do not throw her around on a crowded dancefloor or spin her too much if she is dizzy or tired, but also be active is she is looking to move around. 

As the lead, he will need to pick a level of closeness that is engaging and not too standoffish, but also not too close as to be uncomfortable. Maybe she would like to be to be the flashy centre of attention, or maybe she would prefer to focus on connection. Whether he does this correctly or not is entirely in the subjective view of his dance partner. 

Thankfully there isn’t an independent ‘dancing authority’ that regulates dancing. The male lead cannot appeal that ‘he was doing the right things and therefore she should have enjoyed herself.’ She, because she is being led, is, in the end, the sole arbiter of whether he is doing a good job.  

A reasonable description of what the man’s role is in Salsa dance is to make her look and feel fantastic by her own estimation

The focus of the dancing is actually his partner and not the man. Even though he is ‘leading’ through signals, he must also, in turn, be reading from her the far more complex amounts of information, such as how easy the moves were for her; whether she had to cut moves short so as not to bump someone, or whether she wants to move closer together or farther apart from her dance partner and , of course, how much she appears to be enjoying herself! She is actually leading the man with her subtle cues! 

Indeed her control extends further. She has the ability to accept or reject the man. If he is doing a terrible job on the dancefloor, she could stop and say that she doesn’t want to dance any more because she is tired, uncomfortable, or dizzy. The lady could accept or reject the man when he asks to dance or even cut a dance short. And even when the man asks her to dance in the first place, it is usually only after she has made eye contact, inviting the man to ask her. 

If I had to decide, I would say that in the partnership, it is actually the lady is in charge of the dancing.  

More accurately, though, I think that there are two separate roles that are being performed and that neither is ‘in charge’ of the other. They are both entwined in the dance for their mutual benefit.


Robots and Subtlety 

This brings us to robots. Because the first step in creating robots that mimic what humans do is to model the interactions that humans have. And to model something, we need to have observable data points. The most readily observable data in salsa dancing is the leading movements that the man gives that the lady would follow. An engineer designing a pair of salsa dancing robots would start by teaching the ‘man robot’ a bunch of moves to lead and the ‘lady robot’ the ability to follow them. But such a pair of robots would be a pale imitation of the complex interaction that is actually going on.

In some current research, robots are being inspired by salsa dancing so that they can understand humans and learn to pick up leads from the humans that the robots follow.


This is really cool, but a long way off of the interaction humans do naturally and involves a clear leader (the human) and a clear follower (the robot) as compared with mutual leading as in dancing.

Robots that do seem to pick up this type of subtlety are probably more Wizard of Oz than robot.

Alibaba’s Salsa Dancing Robot @ CES 2015 

Besides dancing, there are many fields in which we attempt to take a very complex set of human interactions, and from the observable surface, data reduce it to an algorithm. In social ‘sciences’ such as psychology, questionnaires will be used to ‘understand’ subjects’ tastes and preferences, a rough form of algorithm. 

However, only about 40% of psychology papers are reproducible! ( See: Over half of psychology studies fail reproducibility test . Largest replication study to date casts doubt on many published positive results . Nature ) Some lower proportion than that gives us insight into humans.  

Machine learning doesn’t seek to explain hypothesis and instead ‘just works.’ See my article Explainable AI Is All The Rage At Legal Technology Conferences Currently . It is engineering rather than science. That is why social media has been so successful because its algorithms don’t try to understand why you react to #fitspo, cat pictures, and stories about Donald Trump, they just put them in front of you the more you engage with them. 

But these algorithms don’t understand non-linear distributions and so cannot predict when people will have too much and want something new. That is why Facebook has been declining the past couple of years. People just get bored and switch off.

 Robots vs Humans

I think that there are a number of things that are best done by humans, one of which is empathy.

 Artificial Intelligence and Law – Arbitration In a Changing World Asia Conference 2017 ( (see here at 8:08)

One of the things that humans do best is to understand other humans. I do not think that robots will ever exceed human-level ability in empathy. This does not mean that no robot will ever be more empathic that any human, there are a lot of un-empathic humans! (There are also a lot of bad human dancers). But a typical human is much better at empathy than even the most advanced robot.

This should not discourage us from trying to teach robots empathy (or dancing), but instead, we should pick the right tool for the job, and when we build systems, we should rely on humans to do the pieces that they do best.