Robots “Learn” To Overcome Injury

Though they have become increasingly advanced, there remains a major difference between most robots and their animal counterparts: the ability to adapt to unforeseen circumstances. Three-legged dogs can learn to run again, but broken robots require human repairs. Now, researchers have reported a major breakthrough, a robot that can suffer a broken limb and independently establish a new gait to work around the damaged area.


Three-legged dog: like the dog shown here, new robots are able to re-evalualte their own motion to compensate for injury.Three-legged dog: like the dog shown here, new robots are able to re-evalualte their own motion to compensate for injury.

Typically, engineers have attempted to approach this problem by programming a series of contingencies into the robot software. Though this can be effective, it has several obvious downsides. First, the robot must be self-diagnosing, which requires additional software and hardware (sensors) and is not always reliable. More importantly though, following diagnosis, the robot must then select from the series of pre-programmed contingencies. The suitability of the contingency possibilities thus relies entirely on the programmer having sufficient foresight into the nature of the damage. The alternative is to learn the way animals do – by trial and error. While, this method does not require any assumptions by the programmer, and is preferable from a manufacturing and operational point of view, it is impractically slow even with small search spaces.

The newly reported method, published this year in the highly regarded journal Nature, is termed “intelligent trial-and-error”. In this system, prior to deployment, the robot “learns” all of its possible behaviors and creates a performance map of each behavior and its value. While this initial step is slow, it only has to be performed one time at the robot’s “birth”. If it is then damaged in the field, it need only tap into this stored performance knowledge to quickly adapt to its new circumstances. To do this, different types of behavior are rapidly tested in an order determined from the pre-established performance map. As the tests are carried out, the robot learns and adjusts the map until it finds what is predicted to be the best possible scenario at which point trial-and-error ends and motion resumes.

In a series of tests, the scientists report that only two minutes are required for the robots to resume motion following damage. A six-legged robot was tested and found to adapt successfully to five different types of breaks. Similarly, a robotic arm could respond to fourteen different joint fractures.

 

There will be tremendous demand for these robots in situations where human intervention is impossible, expensive, or dangerous. An obvious example is of course those used in space exploration, like the Mars rover. Other possibilities include deep sea diving or battlefield-deployed robots. Aside from just “healing” injured robots, the same algorithm will likewise be useful for fully functional robots encountering new obstacles like changes in terrain and other surroundings. Interestingly, this research may also have some benefit in the reverse direction – that is, developing a greater understanding of animal adaptive behavior through analyzing their mechanical counterparts.

What remains to be seen is how many humans will respond to the notion of adaptable robotics. With the inundation of science fiction movies featuring robot armies and artificial intelligence gone awry, it is easy to see how this style of research could generate concern. However, the potential contribution to scientific knowledge, and thus human advancement, from these autonomous robots willing to explore the places too distant or dangerous for us to venture should not be underestimated.

Via Nature.