But when I first heard that NASA was planning to land a solar-powered AI on Mars, I was just astonished. As advanced as NASA’s technology might be, I could not believe this technology was advanced enough to be talking about anything quite like this — and to be perfectly honest, I’m a little skeptic.

Honda's Ashimo

Many of the algorithms we use today for artificial intelligence applications are NP-complete: that is to say, they belong to a class of problems for which we don’t know an efficient solving algorithm just yet, and their typical solving time (and computational power required) increases exponentially or binomially relative to the number of inputs provided.

Since aerospace applications such as NASA’s would need to operate on very short response times, and given the very limited power that can be provided by solar power (even on a planet without an atmosphere), I’m curious to know just what kind of complex decisions this kind of artificial intelligence would be able to make on its own and without human help.

While NASA scientists can probably do without real-time response and let their robots crunch the numbers for a relatively long amount of time, they still need to deal with the huge power consumption AI algorithms entail. Our very best solar panel technology is only 28% efficient, which certainly doesn’t sound like enough to allow for a decent AI to compute anything useful.

This is from a NASA article from some time ago:

NASA is planning to add a strong dose of artificial intelligence (AI) to planetary rovers to make them much more self-reliant, capable of making basic decisions during a mission.

Scientists at NASA Ames Research Center, in the heart of California’s Silicon Valley, are developing very complex AI software that enables a higher level of robotic intelligence.

Later on, the article makes the very valid point that keep instructing our robots from Earth is definitely not the way to go, given the long reaction times that ultimately don’t allow us to exploit these multi-million dollar machines to their full potential. And the waste is going to increase even more with time, once we reach destinations further and further away from Earth.

Developing an AI so sophisticated and efficient that would be able to make the robot completely autonomous would be a huge step forward. Alien Planet, an interesting documentary simulating a mission to a planet light-years away in search for extraterrestrial life, might give you a better idea of what we’re talking about. If mankind is ever to send spacecrafts to distant solar systems, then we really have no choice: the unmanned explorers must be able to take decisions on their own, without human help.

Mars Phoenix - Artist's Concept

One of the leading companies devoted to developing breakthrough AI technology is Novamente, which is pioneering a new approach to the problem of testing and developing what they define as an “artificial general intelligence” — something capable of deduction, abstraction and other kinds of human-like reasoning.

Their approach is to test their AI (which is, by the way, being developed in C++) in virtual words such as Second Life rather than in the real world, in order to cut production costs and not to waste time worrying about countless sensor- and motion-related issues instead of concentrating on the actual problem.

On their site, you can watch videos that demonstrate their work. If you’re into computer science and don’t shy away from a bit of technical talk, their papers section is also quite interesting. What will first come to mind watching videos of the AI in action, though, is that the response time is currently very far from being real-time even for relatively simple tasks, and a powerful computer cluster is needed to keep those times within barely acceptable boundaries.

So, how can NASA be so confident that they will be able to develop AIs capable of elaborating such complex algorithms, including figuring out how to let robots decide where and when to move, and power them with something as inefficient as solar panels?

The answer is simple: according Moore’s law, the computational power of our microchips is increasing exponentially with time. Inevitably, there will come a time when we will be able to run algorithms that now we believe overly complex in an infinitesimal fraction of time. Then, we will have the potential to develop more and more intelligent machines that make good decisions faster and faster.

The real question is: when?