Bionic Handling Assistant

The Bionic Handling Assistant (BHA) is a mature representative from the field of soft  robots but a reliable control of this robot is inherently difficult. We believe that such issues can be principally tackled by application of learning methods. We thereby provide the first functioning control concept for this challenging robot platform.


Soft robots attract an increasing number of researchers, who are are driven by various motivations. The likely most significant treat is the “understanding by building” approach in biorobotics, which strives to complement biological mechanisms with technical implementations mimicing their functionality. From the different perspective of human-robot interaction, for instance in co-working scenarios, fully compliant soft robots are of great interest, because they allow for intrinsically safe and natural interaction. This is unlike traditional rigid-link robotics with classical actuation and stiff control that is confined to be  separated from humans in cages. Despite this obvious advantage of soft robots, the downside of a biologically inspired design is often that hardly any analytic models are  available for control, which in turn is prohibitive towards applications. It is thus often a  fundamental challenge to achieve a degree of practical functionality that is similar to  classical rigid robots. The question is, how to address the essential issues related to  kinematics, dynamics, and control on soft robots? We apply machine learning to address these problems. We demonstrate the power of learning on the BHA, while we believe that these approaches are valid beyond the particular mechanism and could foster breakthroughs also in other domains and soft robotics platforms.


The Bionic Handling Assistant (BHA) is a very mature representative from the field of soft  robots [1]. It has been designed by Festo as a robotic pendant to an elephant trunk, which has gathered strong interest because it belongs to a new class of soft and lightweight robots based on 3D-rapid manufacturing. It is pneumatically actuated and with several  segments of continuous parallel components which are operated at low pressures and  thus inherently safe. At the same time, no automatic control for the BHA has been introduced so far. The major reason is most likely that the BHA comprises substantial  challenges for any classical control scheme including high dimensionality and redundancy, very slow actuator dynamics, active compliance control mode. restrictive and unknown actuation ranges, and non-stationary system behavior due to friction and visco-elasticity.

We will show how learning can be the essential tool to enable control of the BHA in several respects and on several levels.
We thereby provide the first operational control concept for this challenging robot platform based on an advanced multi-level control architecture [8]. The central contributions of our work are multi-faceted:

  • On a lower level, we tackle the problem of slow actuator dynamics, which allow only for very slow and unsatisfactory feedback control, while a precise model-based inverse dynamics control is clearly infeasible on the BHA. We therefore first describe a decomposed model of the inverse dynamics, which considers only equilibria of the robot’s dynamics. This simplification practically enables learning by avoiding too many degrees of freedom and high-dimensional state spaces of such robots. We demonstrate on the BHA that such an inverse equilibrium model can be learned and effectively exploited in a feedforward-feedback-controller for quick and agile control. In a second step, this control scheme is extended to an active compliance control mode, which is now on the user-interaction level. It allows, similar to classical gravitation compensation, to freely move the robot based on the implicit knowledge of gravitational and mechanical forces encoded in the learned equilibrium model. This enables kinesthetic teaching of positions and tasks to the BHA in human-robot interaction. However, to go beyond simple reproduction of positions from teach-in, a full inverse kinematics model is necessary. See [2,3] for details.
  • P { margin-bottom: 0in; }A:link { } We are also able to enable the learning of reaching skills on the BHA, i.e. to move the end-effector of the robot towards some desired position by changing the robot’s posture. Successful control of such tasks can be well understood with the notion of internal models that are learned in an exploratory manner. Once internal models are established for a certain task, a forward model predicts the consequence of a motor command, while an inverse model suggests a motor command necessary to achieve a desired outcome. But learning from scratch is difficult. In particular, if the motor system is complex and subject to non-stationary behavior. The key idea is thus to apply an exploration paradigm called “Goal Babbling” to learn the inverse kinematics very similar to infants that attempt goal- directed actions already days after birth and thus “learn to reach by trying to reach”. See [4,5] for details.
  • This strategy is highly beneficial for high-dimensional motor systems that are also highly redundant: Tasks in sensorimotor learning are typically much lower-dimensional than the motor systems themselves. Reaching can be done in an infinite number of ways, because human bodies as well as modern robotic systems have more degrees of freedom than necessary to solve the task. This is shown in theoretical works that strongly support the usefulness of the goal babbling approach. See [8,9] for details.

  • We also evaluated the use of continuum kinematics with constant curvature as a kinematic model for Festo’s BHA. A new, elegant, and parameterless method to deal with geometric singularities in stretched positions is introduced and the stability of the method is shown with numeric simulations. The model provides a practical, and highly efficient tool for the simulation and experimentation with continuum robots. See [6] for details.

  • Software abstractions of the existing Robot Control Interface (RCI) and the Compliant Control Architecture (CCA) are investigated for the BHA from a software modeling and software architectural perspective. We focus on three different challenges: the first challenge is to enable reasonable and hierarchical semantic abstractions of the robot. The second challenge is to develop hardware I/O abstractions for the prototypical and heterogeneous technical setup. The third challenge is to realize this in a flexible and reusable manner. See [7] for details.

In total, our methods address a learned pseudo inverse dynamics control, a high-level control mode to enable compliant human-robot interaction and the efficient learning of direct inverse models. Non of the results is exclusively tailored to the BHA, we strongly believe that similar approaches are also useful for other soft-robot platforms like worm-like instruments for minimal-invasive surgeries with remote actuation. They may thus pave the way for actual application of soft robots in terms of general manipulation tasks.


YouTube Material:


The success of Goal Babbling on the BHA is shown in the following video:


The kinemtics simulation of the BHA is shown here:



  1. A. Grzesiak, R. Becker, and A. Verl. “The Bionic Handling Assistant - A Success Story of Additive Manufacturing”. Assembly Automation, vol. 31, no. 4, pp. 329 - 333, 2011.

  2. Neumann, K., M. Rolf, and J. J. Steil, "Reliable Integration of Continuous Constraints into Extreme Learning Machines", Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 21, no. supp02, Singapur, pp. 35-50, 12/2013.

  3. Queißer, J. F., K. Neumann, M. Rolf, F. R. Reinhart, and J. J. Steil, "Active Compliant Control for a Pneumatic Soft Robot", IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Chicago, IL, Submitted.

  4. Rolf, M., J. J. Steil, and M. Gienger, "Goal babbling permits direct learning of inverse Kinematics", IEEE Trans. Autonomous Mental Development, vol. 2, no. 3, pp. 216 - 229 , 09/2010.

  5. Rolf, M., and J. J. Steil, "Efficient exploratory learning of inverse kinematics on a bionic elephant trunk", IEEE Trans. Neural Networks and Learning Systems , In Press.

  6. Rolf, M., and J. J. Steil, "Constant curvature continuum kinematics as fast approximate model for the Bionic Handling Assistant", IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS), Vilamoura, Portugal, IEEE, pp. 3440-3446, 10/2012.

  7. Nordmann, A., M. Rolf, and S. Wrede, "Software Abstractions for Simulation and Control of a Continuum Robot", Simulation, Modeling, and Programming for Autonomous Robots, vol. 7628, Tsukuba, Japan, Springer, pp. 113-124, 11/2012

  8. Rolf, M., K. Neumann, J. F. Queißer, F. R. Reinhart, A. Nordmann, and J. J. Steil, "A Multi-Level Control Architecture for the Bionic Handling Assistant", Advanced Robotics, DOI:10.1080/01691864.2015.1037793, 2015



Prof. Jochen Steil

M. Sc. J. Queisser