Robots that can adapt like natural animals

 

Animals are beings that naturally can adapt to any type of injuries while current robots don’t have the ability to create compensatory behavior when damaged: they are either limited to their manufacturer design or need to many hours to search for optimal compensatory behaviors. Antoine Cully, Jeff Clune and Jean-Baptiste Mouret discovered an intelligent trial and error algorithm that gives robots the ability to adapt to damage in less than 2 minutes, thanks to intuitions that they develop before their mission and experiments that they conduct to validate or invalidate after the inflicted damage. The result is a never seen before process that adapts to a variety of injuries, including damaged, broken, and missing legs. This new discovery will make possible the development of more robust, effective, autonomous robots and suggests principles that animals may use to adapt. 

Current self-repair robots have two phases: self-diagnosis and then selecting the best, pre-designed contingency plan. Such robots are expensive to manufacture due to high cost hardware that involves advanced sensors because robot engineers cannot foresee every situation: this way of viewing the problem often fails either because the diagnosis is incorrect or because an appropriate contingency plan is not provided. 

Injured animals respond differently: they learn by trial and error how to compensate for damage. Trial-and-error learning algorithms could allow machines to creatively discover compensatory behaviors. Recovery from damage would be more practical and effective if robots could adapt as creatively and quickly as animals. 

Gaussian process model captures all these ideas, witch approximates the performance function using the already acquired data, and a Bayesian optimization procedure, which exploits this model to find the maximum of the performance function. The robot selects witch behaviors to test by maximizing the points from witch the performance is uncertain and exploits the points from witch the performance is expected to be high. The selected behavior is tested on the physical robot and the actual performance is recorded. The algorithm updates the expected performance of the tested behavior and lowers the uncertainty about it. 

An example is that after a damage occurs the robot is unable to walk straight and damage recovery via Intelligent Trial and Error begins. From an automatically generated behavior repertoire robots tests different types of behaviors. After each test, the robot updates its predictions of which behaviors will perform well despite the damage. This select/test/update loop is repeated until a tested behavior on the physical robot performs better than 90% of the best predicted performance in the repertoire, a value that can decrease with each test. This way, the robot rapidly discovers an effective compensatory behavior. 

A parallel is that the simulator and Gaussian process components of Intelligent Trial and Error are two forms of predictive models, which are known to exist in animals. 

  

    Consulted during research: 

Robots that can adapt like natural animals by Antoine Cully, Jeff Clune and Jean-Baptiste Mouret