[video_player id=鈥滵agrnwEb鈥� access_level=鈥漞veryone鈥漖
A robotic elephant trunk that uses artificial intelligence to mimic some aspects of brains could lead to snake-like machines that can roam and adapt to new tasks.
Sebastian Otte at the University of Tubingen in Germany and his colleagues created a 3D-printed robot trunk from segments that each include several motors driving gears that tilt up to 40 degrees in two axes. The trunk can bend, but also elongate or shorten.
Advertisement
The team created a trunk with 10 segments, but they say the length could be doubled with more powerful motors. In tests they found that the brain-like AI could direct the tip of the 80-centimetre robotic trunk to a target to within less than a centimetre.
Controlling a robot with this many degrees of freedom is so tricky that traditional computer programming quickly becomes very complex. Instead, robots operated by AI typically use a neural network which mimics the operation of a brain with large networks of neurons connected by synapses. To control the trunk, examples of the various motor inputs needed to move it in certain ways are fed to the AI during a training stage, and an algorithm to control it is produced.
The team used this robot as a proof of concept for a relatively new type of algorithm that offers a vast improvement in efficiency, called a spiking neural network. These work like real brains, in that certain inputs cause a chain reaction of firing synapses, and they require orders of magnitude less computational power and energy.
Typical neural networks are computing all the time, even if there鈥檚 no activity, says Otte. 鈥淚t鈥檚 redundant and it鈥檚 not necessary, and this is different in spiking neural networks. When there are no spikes there鈥檚 no computation,鈥� he says.
A lightweight robot with a relatively basic computer could run a spiking neural network and continue to train itself when put to work, potentially becoming more efficient or learning new tasks. Robots that are trained with an external computer typically have algorithms that can鈥檛 adapt in real time.
鈥淥ur dream is that we can do this in a continuous learning set-up where the robot starts without any knowledge and then tries to reach goals, and while it does this it generates its own learning examples,鈥� says Otte. 鈥淵ou would ship out a pre-trained network that knows something, but it tunes itself for the particular application.鈥�
The researchers believe that it might be possible to make snake-like robots where the trunk isn鈥檛 tethered, but can roam at will. This might prove useful in search and rescue operations as it could weave into small spaces.
Reference: arXiv,
