Sticks Transformed Into Mobile Robots Using Deep Reinforcement Learning

While walking robots generally rely on carefully designed mechanical structures, researchers from the University of Tokyo and Preferred Networks have been able to build them out of tree branches and other unusual materials. No, not branches that have been carefully milled to the proper dimensions, but branches, as in something that you literally pick up in the forest.

Such an astounding feat of robotics is accomplished with the help of machine learning — specifically, “deep reinforcement learning” — with the sticks first weighed and 3D-scanned. These pieces are then virtually put together in a simulation, along with servo actuator elements, allowing the computer to choose which gaits result in the best movement. There is also reportedly some hand-tuning involved to account for movements that might break the robot, or that appear a little off to the project’s human overlords.

The simulated stick-bots were then translated into the real world, controlled using an Arduino Mega, and powered by Kondo KRS-2572HV servo motors with a separate driver and power supply. Notably, the device uses a tether here, apparently to facilitate control and weight savings. We are, after all, talking about sticks literally picked up off of the ground!

Although the practical applications for such a robot might seem limited (though, per the video, at least three have been created), being able to locally source much of the building materials for a ‘bot could have some important benefits. On the other hand, perhaps such leaning algorithms could be used to simply make good designs even better.

[h/t: IEEE Spectrum]