Kuroki – Machine Learning on a Robotis Mini

In my research I have the opportunity to learn about machine learning (ML). Knowing ML is a fundamental to understand the capabilities and limitations of current robotics. If you are new to the field, I fully recommend the machine learning course by Andrew Ng and the probabilistic graphical models specialization by Daphne Koller, both on Coursera.

While I directly applied and developed some ML techniques in my research (i.e., robust estimators, probabilistic models, reinforcement learning, …), I did not “get my hands dirty” with some others. I believe the best way to learn something is to use it to solve a problem, so I looked for an excuse to learn more about these other ML techniques I didn’t use before.

My main goal is to get more practical experience on deep neural networks (DNNs) and reinforcement learning (RL), two techniques that are receiving increasing attention on robotics due to the awesome results they have delivered in games (for example, Alpha Go, or agents playing Atari) but also in manipulation, locomotion, grasping… DNNs and RL are powerful horses for robotics but, as any other tool, they present some strengths and some limitations. I want to explore those!

And the excuse to explore DNNs and RL came last Christmas! Its name is Kuroki, it is a Robotis Mini unit and it is amazing! It was not a surprise: my family wanted to give me something I would really be happy to use and I came with the idea. After searching for a while I realized that: 1) robots are still pretty expensive when it comes to hobbiests, 2) Robotis Mini is the best robotic platform to acquire if you want to focus on the “brain” of the robot and not on the hardware, which is my main interest and skill-set.

You can get the robot for 300€ and it can be programmed/used at different levels of complexity: you can play with it using an app on your smart phone and commanding some predefined motions (nice for demos for your family and friends); you can modify those motions and create sequences of motions using a Robotis software with very nice GUI (only for Windows and Mac, though); and finally, and most importantly to me, you can program the controller in a C-like language. The controller is a OpenCM 9.04c which can be programmed very similarly to how you would program an Arduino.

However, I encounter rather quickly that the C interface is rather limited for any interesting and complex tasks that can make use of DNNs and RL for control. In next posts I will explain how to modify the Robotis Mini to control it directly from your computer and be able to use the computer as perception, planning and control unit.

Enjoy!