The Bionic Handling Assistant

The Bionic Handling Assistant (BHA) is a new robot platform inspired by elephant trunks which has been developed by Festo. The robot is pneumatically actuated and made almost completely out of elastic (!) polyamide which makes it very flexible and lightweight. Yet, its bionic design thoroughly challenges standard methods to deal with robot systems, which concerns control, simulation, and software architecture.

Bionic Learning

An important step towards leveraging the robot's potential, e.g. for physical human-robot interaction, is to develop an inverse kinematics controller. The robot should be able to get its gripper on some object in order to manipulate it. Sounds easy, but isn't... at least not if one searches in the standard robotics toolkits. Yet, it is a nice use case for Goal Babbling, and one that actually matters. The video shows how the method (based on my ICDL 2011 paper) learns to control the robot. It also shows how the method can deal with possible hardware defects, and how teaching in pHRI can be leveraged on the fly. Results on the learning method, performance, and a proposed integration with feedback control are published in this paper.
See also:
The learning on a kinematic level builds on top of a postural controller that is likewise powered by machine learning. Postural control for the BHA means to control the length of each actuator by means of pressure inside it. In principle this can be done with standard PID control, but only very badly and slowly. The sensors are too noisy and the delays in the pressure control and particular the delays between pressure and actual motion are way too long for such simple feedback control. The solution is to use a learned feedforward controller in addition to feedback control. Getting enough (and good!) training data for such a controller is hard. Yet, on the level of postural control we can fill up this lack of good training data with prior knowledge - at least to some extend - and we can be use it in order to constrain the learning. A novel, and very powerful method to combine training data and prior knowledge has been developed by Klaus Neumann. We describe the method and results for the BHA learning in this paper.

Kinematics Simulation

The BHA is really hard to model... exactly. Yet, there is room for approximate models being very useful. On a kinematic level a (forward) model predicts how the gripper of the robot moves, given a certain movement of each actuator. There have been various models suggested in literature to describe the elastic motion of continuum robots (like the BHA). They all range somewhere between making very, very simplifying assumptions (sometimes much too simple to describe reality) and using absolutely sophisticated physics simulations (which are computationally very expensive). Surprisingly, we found that simple models can perform quite well (though not perfect of course) for the BHA. The method and some experiments showing the acuity on the BHA have been demonstrated in this paper. The software for this model is open source, so you can download and use it. For our purposes this piece of software has done an excellent job in various contexts such as visualization and prediction.
See also:

Flexible software architectures

Once again: this robot is challenging! Also to deal with it from a software perspective. Typical software abstractions for robotics like "one actuator" are meaningless for the BHA's parallel actuation design. The interfaces to the hardware are heterogeneous, to say the least. Plus, we needed a lot of flexibility on the lower level software layers, while already building on them. Puh...
For such occasions it is really good to have people who think about software engineering. Arne Nordmann is developing a software framework to deal with such (and more) challenges in robotics, as part of the AMARSi European Union project. This paper studies how to use it for modelling the BHA.