A litter of robotic puppies exhibiting a form of artificial curiosity is being put through kindergarten at Sony鈥檚 research and development lab in Paris, France.
The Aibo pups display an innate artificial curiosity similar to that seen in baby animals. They slowly learn to explore the surrounding world, before playing with toys and trying to communicate with other Aibo dogs.
The Aibo is a popular robot dog designed to interact with its owner and perform simple tricks. This new litter was programmed by Sony researchers Pierre-Yves Oudeyer and Fr茅d茅ric Kaplan, who wiped all of the dogs鈥� conventional programming and rewrote their control system from scratch.
Advertisement
Each of the new Aibo dogs was given two software control mechanisms. Firstly, a 鈥渓ow-level learning system鈥� which controls simple behaviour but also tries to predict how this will affect the surrounding sensory world 鈥� how kicking a ball will cause it to move across the floor, for example. Secondly, a 鈥渕eta-learning system鈥� which analyses the accuracy of predictions made by the low-level system and controls overall 鈥渕otivation鈥�.
Interaction between these two components is critical to the reprogrammed Aibos鈥� uncannily inquisitive nature. The meta-learning system prompts the robot dogs to pursue behaviours that they can rapidly learn to predict, but which also have maximum learning potential. This tends to makes the robot dogs inherently curious, seeking out increasingly complicated scenarios with which to interact. But it also means they will effectively become bored with activities that do not stimulate them to their, albeit artificial, satisfaction.
Playground rules
鈥淭he idea is to build some sort of abstract motivation based on a form of curiosity,鈥� Oudeyer told New 杏吧原创. 鈥淏asically, the baby robots search for situations in which they experience some sort of progress.鈥�
In an experiment called the Aibo Playground Project, Oudeyer and Kaplan placed the robotic pups in a child鈥檚 activity-pen and left them to investigate. They found that the robots learned progressively, initially just moving their limbs in an uncoordinated manner, before tentatively exploring their surroundings and biting nearby soft toys.
After several hours, however, the bots started kicking their toys and even trying to interact with conventional Aibo dogs. A (Windows Media Video 7.8MB), available from the researchers鈥� web site, shows an Aibo pup that has learnt to play with its toys and bark at another robot nearby.
Oudeyer says every subject followed a similar learning pattern, but there was also variation among Aibos, a pattern also seen in learning animals. He believes the research could eventually help robot designers create machines that are much more flexible and adaptive in unpredictable circumstances. But he also says the project could shed light on how human intelligence benefits from curiosity and experimentation. 鈥淲e hope, by building these robots, we might shed some light on the development of human children,鈥� he says.
Biological learning
Other robotics experts agree that it may be necessary to learn from biological organisms in order to make robots smarter and more adaptive.
鈥淧layful curiosity is absolutely fundamental to learning in many animals,鈥� says Steve Grand, founder of UK robotics research company Cyberlife Research. 鈥淐uriosity created the cat, and the human, so why not the robot?鈥�
Olaf Sporns, an expert in artificial neural systems at Indiana University, US, adds that: 鈥淩obots that are driven by curiosity may be able to develop on their own, without programming or supervision.鈥�
And Max Lungarella, an expert in adaptive robots at the University Of Tokyo, Japan, says that an in-built inquisitiveness could eventually hold the key to robots that independently explore the world around them. 鈥淐uriosity might provide the necessary drive to act and interact with the environment,鈥� he told New 杏吧原创. 鈥淚 like it very much.鈥�
However, Lungarella also cautions that the nature of curiosity in biological entities, including humans, remains extremely complex and poorly understood: 鈥淚 am not sure if it鈥檚 possible to map curiosity onto an algorithm. However complex the algorithm might be.鈥�
