麻豆传媒

THE pearly white humanoid watches placidly as the woman moves a toy brick sitting on the table. Inside, iCub鈥檚 imagination is running wild.

The robot is being tested for its ability to track the mental states of others. Known as theory of mind this gives humans many sophisticated traits, including empathy and deception Robots have demonstrated theory of mind before but iCub is different. Last week, at the in London, researchers revealed that it is the first robot to acquire theory of mind without specific programming. 鈥淭his all emerged,鈥� says , leader of the research team.

Dominey is one of a band of roboticists who are showing that building with basic biological machinery 鈥� instead of ever-more complex algorithms 鈥� can endow robots with lifelike characteristics. 鈥淲e can directly take advantage of the evolutionary lessons of nature,鈥� says biologist . 鈥淲e are not forced to rely on the conjecture of engineers.鈥�

Many successes in artificial intelligence are largely due to having abandoned the attempt to model human thinking (see鈥�Not like us: Artificial minds we can鈥檛 understand鈥�). But several roboticists believe that for robots to acquire some complex traits found in the animal world 鈥� like social skills, or the ability of insects to react to changes in airflow during flight 鈥� machines still sorely need biology. These biomimetic machines, in turn, are helping biologists hone their understanding of animals (see 鈥�Beebot explains the bee鈥�).

Dominey鈥檚 iCub is a case in point. It has an accurate model of human autobiographical memory. Like ours, this is split into an episodic memory for recording specific events, and a semantic memory, which finds patterns in the events and turns them into rules, or knowledge. Dominey鈥檚 team at the Robot Cognition Laboratory at INSERM, the French national medical research agency, in Lyon, France, found that this gave iCub a more natural form of learning. For instance, by recording the coordinates of objects placed on a table in front of it, iCub learned for itself that the word 鈥渓eft鈥� didn鈥檛 refer to coordinates, but rather to the relative positions of objects.

Theory of mind was an accident that emerged from the next stage of experiments. Psychologists believe that humans use a simulated internal self to learn. Imagine you are about to grasp a glass. Internally, a part of your brain is predicting the amount of force needed to hold it. If, in the event, you break the glass, your brain makes a note of the difference between prediction and reality, and adjusts its knowledge of the world accordingly.

To mimic this, Dominey鈥檚 team gave iCub a simulated internal self. Every time the robot executes an action, like moving an object, it instructs its own arm to move and records the result. In parallel, through an identical programme, it gives the same instructions to simulated iCub. Instead of carrying out the action, the simulated iCub uses its knowledge from past events to predict the outcome. Any differences between the real outcome and the simulated one are a sign that iCub鈥檚 knowledge needs updating.

Crucially, because this gave iCub two versions of itself, the researchers realised that they had inadvertently created the set-up for it to understand the mental state of others. All they had to do was link simulated iCub to another individual, rather than itself, which they did by instructing simulated iCub to stop updating its memory if that person wasn鈥檛 present.

Free mind

Humans typically acquire theory of mind around the age of 3 or 4. Whether they鈥檝e reached this mental milestone can be determined by the Sally-Anne test. In it, a child is shown two characters, Sally and Anne. Sally puts her ball in a basket and leaves, then Anne moves Sally鈥檚 ball into a box. The child is asked where Sally will look for her ball when she returns. Children who have theory of mind will correctly say that Sally will look in the basket, even though they know the ball is now in the box.

When facing the test, iCub passed by comparing its own record of events with the record belonging to simulated iCub. 鈥淲e get theory of mind for free,鈥� says Dominey. His student, Gr茅goire Pointeau, who presented the results, says they could be used to make robots that can anticipate the needs of others.

Other teams are also watching lifelike behaviour emerge from their machines. To give robots powers we don鈥檛 have, such as flying, Ayers and his colleagues are copying the electrical activity of insect nervous systems, with the aim of creating an artificial bee that could pollinate plants.

鈥淭o give robots powers we don鈥檛 have, such as flying, we are copying insect nervous systems鈥�

Previous attempts to imitate animal movement have relied on building equivalents of specific abilities, like magnetic compasses and air velocity sensors, but such robots are still controlled by computer algorithms. 鈥淭he problem with algorithms is that you have to anticipate every possible situation and have a determined escape strategy for each,鈥� says Ayers. That鈥檚 a pain when you鈥檙e doing something as prone to variation as flying.

His team鈥檚 secret is circuit boards that produce chaotic electrical signals 鈥� much as real neurons are thought to 鈥� allowing a multitude of possible solutions to be explored on the fly. 鈥淐haos allows you to explore your full parameter space until you find a solution that allows you to escape,鈥� says Ayers.

The approach seems to be working. At the Living Machines conference, he showed how his team had controlled the flight of a small toy helicopter with a synthetic nervous system. Much like a bee鈥檚 waggle dance, which tells other bees where food is, the team transmitted a distance and a direction to the helicopter, which it successfully followed. 鈥淭o my knowledge this is the first time someone has controlled an aerial vehicle purely from biological knowledge of networks,鈥� says Ayers. The next step is to use the same neural control system to pilot RoboBees (see picture). The circuit boards use less power and are lighter than a GPS navigation system would be. The approach could be used to integrate other lifelike sensors, such as an artificial nose, into flying robots.

This new strain of biomimetics isn鈥檛 confined to brains and nervous systems. Later this year, a new journal, Soft Robotics, will be launched. This reflects a trend towards squishy robot bodies that replicate the structure of biological soft tissues and so have the same adaptive properties without being programmed to respond to every eventuality. Editor-in-chief Barry Trimmer , loves the example of caterpillar skin. 鈥淚t has unlimited degrees of freedom but there is no supercomputer,鈥� he says.

This article appeared in print under the headline 鈥淢achines come to life鈥�