Overview : The Robot Studio is a leading designer of open-source humanoid robots based on the same mechanical principles as the human body. This approach produces a style of robot capable of adapting to its environment, with movements that intuitively differ from the rigid motions of a standard robotic system. We aim to create a robust, free hardware platform that can safely perform a wide range of useful tasks alongside humans in everyday real-life situations.

Problem description : With such complex robots it is not possible to use classical methods like the inverse kinematics to drive the arm in space like it can be done for a stiff industrial robotic arm. In this case there is too much non-linearities, too many degrees of freedom with some redundancy and elasticity in the system.
One of the possibilities is to train the robot to learn a specific task using Deep Reinforcement Learning algorithms. But it would be a very tedious long process to do with only one robot prototype because it would need a lot of human assistance and the hardware would probably break before it has learned the task. The way to go around that is to do the training in a simulator, duplicate the experiment a thousands times in parallel on GPU-servers and at the end, transfer the skill to the machine and let it learn from there.

Project : The main goal of this project is to simulate an anthropomimetic robotic arm using OpenSim and to link it to some machine learning libraries like TensorFlow to apply DeepRL.

To have an idea on how it looks and move, here is a video of the robot arm, driven from a gamepad : https://www.youtube.com/watch?v=m1l0T5p3lN0

Any advice or help on how to achieve that is greatly appreciated!

Website : http://www.therobotstudio.com