Robotics and Cyberphysical Systems Lab


Vision-based Furuta pendulum

Picture of a scientist behind a Vision-based Furuta pendulum Copyright: © Bleher / MPI-IS

The vision-based Furuta pendulum is a standardized experimental setup for camera-based observation of dynamical systems. It allows us to learn to control dynamical systems directly from pixels, and can be used for different research purposes.

The furuta pendulum is a challenging system since it is nonlinear, unstable and requires precise and fast control. Second, the standardized camera setup can be added as a high dimensional observer to the angle encoders of the Quanser Qube Servo 2 pendulum. The brightness of the light source can be varied and the Flir BlackFly S camera can be sampled up to a frequency of 522 Hz.

In combination, this setup is used to develop, test, and validate learning algorithms in flexible manner. The Python interface is based on the OpenAI Gym standard and can be used with and without camera-based observations, in a 3D Mujoco simulation and on the hardware setup


Cyber-physical networking testbed

Graph: Cyber physical Systems and multi hop networks Copyright: © Dominik Baumann / MPI-IS

To develop cyber-physical systems (CPS) into reliable and accepted technologies, it is indispensable to test, validate, and benchmark existing solutions in practice. Particularly, it is key that the end-to-end system, including all hardware and software components, is evaluated systematically on physical platforms and real-world wireless networks in scenarios that resemble as close as possible the targeted CPS use cases. Most existing testbeds, however, independently developed in the control and wireless networking communities, study the performance of control designs and wireless protocols in isolation. To overcome these isolated procedures and enable an end-to-end analysis, we built a novel cyber-physical testbed, as illustrated in the figure above. The testbed consists of multiple cart-pole systems that are connected over a multi-hop wireless network. The testbed has first been proposed in [1] and since then extensively been used in our research on wireless networking and control (see our project: Wireless Control for Cyber-physical Systems):



A picture of the Wheelbot Robot Copyright: © Rene Geist / MPI-IS

The abilities to self-erect and maneuver around moving obstacles are crucial steps towards the deployment
of fully autonomous mobile robots acting in uncertain environments. With this perspective
in mind, learning algorithms for robotics must be tested on real-world systems, as hardware raises a
multitude of additional challenges. In the field of machine learning and control, the preferred testbed
for algorithms are either rotorcrafts (quadrocopters), legged robots, or stationary systems (robot arms, pendulums).
While research on such testbeds leads to breakthroughs in learning control and motion planning, analyzing
learning algorithms for naturally unstable non-holonomic systems has been rarely investigated.
With direct representatives such as motorcycles, and airplanes the development of a simple small-scale
testbed is of great significance for the development of learning control algorithms for naturally unstable agile systems.


Other Infrastructure

Picture of Apollo Robot Copyright: © Wolfram Scheible / MPI-IS

We are in the process of setting up additional robots and cyber-physical experiments. Stay tuned.