Two Times Out of Five: A Developer Among Social Robots

As a lover of Wall-E, of Baymax and even of Chappie, it was out of desire and out of the ambition to change the world that I went into social robotics two years ago.

My daily work is designing and writing software for interactions between humans and robots. It goes into welcome desks, into playing with children, into presenting a new product in a shop, and maybe one day into keeping someone company at home. The machines on my desk are called Nao and Pepper.

I have a confession. From what I had been reading online and watching in videos, I believed we were on the edge of the singularity. I have had to lower my expectations.

To put it briefly: forget what you think you know about robots. They are not clever at all. They cannot beat your phone. I would even say they are dumber than it is, because they cannot lean on Google’s or Apple’s intelligence.

They do not understand what I say

I had the chance to meet a lot of them at CES 2017, and I was struck by the growing interest in this industry. New ventures appear every day. In some aisles I could not take a step without running into a robot I had never heard of. I felt like a child on Christmas morning.

The hardware is often very well designed, and Hease Robotics is a good example of that. The software, much less so. Four robots out of five, among the ones I met, did not understand what I said to them. The show’s internet connection kept dropping, speech to text went with it, and nothing was left.

That detail says almost everything. The intelligence is not in the machine standing in front of you, it is in a data centre thousands of kilometres away, and the robot is a thin client with a face. Cut the wire and it goes back to being a plastic shell. Which is all the more frustrating when you know where voice recognition stands today.

The gap between AI and robot intelligence

That leaves the question of why there is such a gap between what AI can do and what a robot does. Several things are at work.

First, deep learning is a discipline that has only just been born. It needs time to grow and to be understood before it can spread into other sectors.

Second, and this is more awkward, we do not always work on what matters. It is very satisfying, as a developer, to announce that you have managed to write a facial hair recognition algorithm. It is a technical success, and it is perfectly useless. In robotics, by contrast, a robot able to detect sadness in a voice, and to react to it, would change something. That is where the real interest begins, commercial interest included.

So let us be patient. I am convinced that good things lie ahead, and it is even likely that within a few years robots will be able to detect and recognise just about every stimulus around them.

Creating is easy, developing is harder

We often hear about the achievements of AI and robotics specialists, and that work deserves credit. What we never see is how they got there.

An algorithm can be marvellous in one specific circumstance and fail completely the moment you move it somewhere else. That is the problem you face as soon as you want to write a new one: it has to work in every situation, not in the one you demonstrated.

Take a concrete case. I can give Pepper the ability to detect any phone coming towards her. It is simple to program, and it is charming: she offers to take a picture with the visitor. Except that two times out of five, the algorithm misreads the scene and the machine embarrasses itself.

To avoid that, you test endlessly. And it takes an eternity to go from prototype to finished product. That is where the time goes, not into the idea.

What the robot does matters as much as what it detects

Let us stay optimistic: AI has a real chance of bringing a great deal to social robotics. But I want to insist on one point, because it is the one most often forgotten.

The keystone of a social robot is not stimulus screening. What the robot does matters as much as what it detects, and that is what has to be solved.

Suppose sadness has been correctly detected. What should the robot do? Sympathise and match the mood, or make the person laugh to comfort them? There are any number of situations of this kind, which seem obvious to understand and are headaches to develop.

That is exactly what we work on at Hoomano. The empathy system built for the Cozmo robot is a good example: its small face, somewhere between Eve and Wall-E, is a wonderful tool for making people want to interact.

Despite everything I have just written, it is a pleasure and a genuine challenge to work in this industry every day. It is completely different from other software sectors: you have to think differently and act differently. The field is new and the opportunities in it are large. Every day brings a new problem, and that is what I love about it.

I can only encourage the curious to come and have a look.