STIFF EU project on enhancing biomorphic agility through variable stiffness - DLR hands - logo by Ian Saunders - artificial arm and hand by TU Delft

STIFF is a research project on enhancing biomorphic agility of robot arms and hands through variable stiffness & elasticity. It is funded by the 7th framework programme of the European Union (grant agreement No: 231576).

Check our 2011 Summer School on Impedance

Institutional Partners
German Aerospace Center (DLR), Germany:
Project coordinator. Responsible for integrating a variable-impedance robotic system in the project. Development of a novel EMG system for human impedance measurements. Integration of human and robotic impedance control approaches.

Technische Universiteit Delft, Netherlands:
Responsible for modelling the human neuromuscular system from muscle to joint level. Developent of time varying system identification and parameter estimation techniques to quantify the model parameters from recorded data using haptic manipulators.

IDSIA, Switzerland:
Responsible for learning high-level task-specific controllers based on reinforcement signals for the flexible variable-impedance robot arm developed by DLR, and for inverse reinforcement learning to extract cost functions in collaboration with UEDIN.

University of Edinburgh, United Kingdom:
Responsible for the development of 'Optimal Feedback Control' based closed loop control paradigms, specifically tailored to redundant and variable impedance actuators. Developing methods to extract cost functions and comparing control policies to evaluate improvement in performance when modulating impedance optimally.

Université Paris Descartes - CNRS, France:
Responsible for studies of impedance control in humans, using a variety of techniques including direct physiologicial measurements (EMG, H-reflex), mathematical modeling and robotic simulation. The main emphasis is 1) to suggest biologically-inspired strategies to be applied to robotics control and 2) to use analogies with robotic devices to better understand human behaviour in terms of impedance.

artificial DLR hand grabs a glass; humanoid robot javelin thrower cartoon by Juergen Schmidhuber



Part 1 (responsible: UPD)

A major objective of the STIFF project is to understand how human beings benefit from stiffness control in different motor behaviours. Insights gained from studying human motor behaviour can then be transferred to artificial systems to see how performance might be improved in robotic systems and to determine the benefits of biomimetic actuators that allow for the control of passive stiffness properties. This part groups together the experimental studies to be carried out with healthy human volunteers. First, better estimates of the time-varying characteristics of human stiffness will be developed. Second, the study of human stiffness behaviour will be extended beyond what is currently known based on artificial laboratory tasks. Finally, we will look at the neural mechanisms underlying stiffness control in humans. These activities will have a close interplay with other Parts; human behaviour will inspire control schemes to be implemented on robotic systems while models of the human musculoskeletal system and implementations on robotics systems will be used to better understand the observed human behaviour. Human (stiffness) performance during the experiments will be quantified as a benchmark system for the robotic performance with and without the controller developed in Part 4.

Part 2 (responsible: DLR)

Building on a variable impedance integrated hand-arm system currently under development at DLR and funded elsewhere, this Part mainly deals with the challenge of (i) using variable impedance actuators and (ii) modelling the control and actuation in this novel robotic hardware. Optimal Control of this actuator system is investigated in detail (without considering optimality of stiffness at this stage) since doing this efficiently in a variable impedance scenario has not been investigated except in toy systems. Various demonstrator robotics tasks will be used to mimic and compare the results of the human impedance measurements in Part 1, including ball throwing.

Part 3 (responsible: TUD)

This Part deals with detailed modelling of the arm-hand musculoskeletal system with biomechanical fidelity. Why? Because we want to use results from the studies of Part 1 in a more quantitative framework that is conducive to optimization studies. The musculoskeletal model can be treated as a physically realistic plant (like a robot!) in which the geometric, actuator and reflexive properties are well-defined, from which the optimization principles and strategy can be gleaned (will be done in Part 4) based on movement study data from Part 1.

Part 4 (responsible: UEDIN)

This Part will first extract key optimality principles used in biological systems by combining data from the human studies (Part 1) in the context of the bio-mechanically realistic human arm models (Part 2). These optimality principles can then be implemented on (a kinematically slightly different) robotic variable stiffness system. This ensures that the mismatch between the exact human arm model and the tendon based robot arm is not an issue since we are extracting the principles. Higher level control goals will be specified through RL methodology followed by low level optimization of redundant joint angles and stiffness parameters based on the OFC framework. Results will be benchmarked with criteria developed in Part 1.

artificial DLR arm and hand; artificial hand squeezes STIFF

Part 1 deals with the challenges of measuring dynamic human arm stiffness during movement tasks and develop novel methodologies to achieve this. The neuro-physiological basis for variable impedance control in human arm is also investigated to achieve deeper insight into the mechanism for such modulations. Part 2 mainly deals with the implementation and control of variable impedance in a ground-breaking, antagonistic robotic hand-arm system based on biomorphic principles. Part 3 is devoted to anatomically realistic biophysical modelling of the human hand-arm system (at the level of muscles and their connectivity) and human grasp force modelling during manipulation tasks.. Exploiting the stiffness estimates from Part 1, parameterized potentials and/or cost functions from human movement plans are approximated in Part 4. These are incorporated in the control theoretic analysis of optimal feedback control in the muscle model from Part 3. The optimization principles gleaned through this analysis is used in turn to control the redundant, variable impedance robot arm. Effectiveness of the control strategy will be tested on various robotic movement case studies that are explored in Part 2 and compared to human movement studies from Part 1.

artificial DLR hand holding a wine bottle