Leo is a prototype that is especially designed for machine learning experiments for humanoid robots. Its boom construction always keeps the hip axis horizontal, thereby making it effecively a 2D robot. The boom also provides power to ensure lengthy learning experiments. Other than that, the robot is fully autonomous. Computing is done on-board using a single board computer with 1.2GHz VIA processor and 1GB of RAM. Leo has 7 joints - three in each leg and one in the shoulder - that are formed by Dynamixel RX-28 servo motors. The Linux OS with Xenomai realtime extension ensures hard real-time control loop characteristics for both conventional and learning controllers.

Leo can stand up autonomously after a fall. Together with the boom construction, that provides power and makes the robot walk in circles, this makes Leo a fully autonomous platform that is especially suited for machine learning experiments.(Download the movie)

Adapting the learning algorithm

The learning process sometimes causes LEO to behave clumpy and to be potentially harmful to itself. LEO's learning algorithm causes its gears to fail four times as fast during the initial learning phase as during the walking phase. The learning algorithm causes tremors which in turn cause high loads in the gear teeth. These loads can be reduced or sometimes prevented by adapting the learning algorithm. The adapted learning algorithm has restrictions on how fast it can change the torques around the joints. The effect of this restriction on the mean time between failure of the gears inside the RX-28 motors has been investigated.


Name author: Erik Schuitema
© 2016 TU Delft