The Moravec Paradox, or Why Robots Are Clumsy at Walking

A calculator handles complex operations and basically spits out results. An artificial intelligence analyzes in minutes what would take a human years. Yet ask any robot to walk down three flights of stairs, and except for extraordinary exceptions, it will never get past the first step. That contradiction is the essence of the Moravec paradox: advanced tasks based on calculation and reason require little processing power, while motor and sensory skills go in the opposite direction.

I figure by now you’ve probably seen a couple of videos of skilled people working at high speed. Wrapping packages, cooking, moving materials, laying bricks, and more. These people’s skill is impressive, and repetition is a fundamental aspect of their development. Now, these and other tasks involve very low-level processes that are unconscious, and we execute them almost perfectly. The real challenge is to transfer these processes to artificial intelligences or systems. Turning a computer into a chess player is relatively easy. But if you want a robot to wash the dishes for you... that’s a very different story. Welcome to the Moravec paradox.

The Moravec paradox: What’s difficult is easy, and what’s easy is difficult

It all started with the work of Hans Moravec, Rodney Brooks, and Marvin Minsky. Moravec states in his 1988 book Mind Children that it is easy to make a computer demonstrate performance on intelligence tests similar to that of an adult human, but it is very difficult or even impossible to give it the abilities a 12-month-old baby has in terms of mobility and perception. Minsky highlighted in his text Society of Mind (1986) that the entire set of low-level processes in a human is highly resistant to reverse engineering, and that “we are not aware” of our best abilities. On the contrary, our attention focuses on simple actions that take more work, not on the complex processes that occur naturally.

Moravec believes the key behind the paradox lies in evolution. All the motor and sensory elements of the human brain are backed by millions of years of experience, and the older a skill is, the more time natural selection has had to optimize it. Of course, this isn’t limited to climbing or running. Judging someone else’s intentions or recognizing a threat are just two examples in an extraordinary package. At the same time, the paradox is a harsh blow to the initial optimism in developing artificial intelligences. At the time, experts believed that by solving complex tasks first, everything was ready to master things like vision, recognition, and common sense. Not only were they wrong, but they’ve been fighting for decades against those “routine” details.

Source: Xataka Ciencia