KI - Künstliche Intelligenz

, Volume 26, Issue 4, pp 407–410 | Cite as

How Rich Motor Skills Empower Robots at Last: Insights and Progress of the AMARSi Project

Projekt

Abstract

Flexible, robust, precise, adaptive, compliant and safe: these are some of the qualities robots must have to interact safely and productively with humans. Yet robots are still nowadays perceived as too rigid, clumsy and not sufficiently adaptive to work efficiently in interaction with people. The AMARSi Project endeavors to design and implement rich motor skills, unique flexibility, compliance and state-of-the-art learning in robots. Inspired by human-recorded motion and learning behavior, similarly versatile and constantly adaptive movements and skills endow robots with singularly human-like motor dynamics and learning. The AMARSi challenge is to integrate novel biological notions, advanced learning algorithms and cutting-edge compliant mechanics in the design of fully-fledged humanoid and quadruped robots with an unprecedented aptitude for integrating in our environments.

Keywords

Adaptive behavior Compliant systems Learning Robotics 

1 Inspiration

Compared to animals and humans, the motor skills of today’s robots are still poor. Their movements are still perceived by people as abrupt, unpredictable or unnatural. In fact, the behavior of robots is often limited to a narrow set of carefully programmed motor patterns that operate a rigid mechanics and display limited adaptation to complex, task-oriented behavioral patterns. On the other hand, the smoothness, efficiency, elegance and safety of movements of humans and other animals are aspects that make the human-to-human or human-to-animal interaction still qualitatively superior to that of any robot.

The AMARSi Integrated Project (Adaptive Modular Architectures for Rich Motor Skills) is a EU funded four-year research project that aims at a qualitative jump towards biological richness of robotic motor skills. Richness is intended as a novel conception of motor primitives as a basis to a large repertoire of motor behavior, ranging across the entire hierarchy from simple periodic and aperiodic motions to complex, task-oriented interaction sequences between a robot and a human caretaker. The AMARSi project proposes a coordinated research effort in different areas (Fig. 1) to implement rich motor skills. The research on rich motor skills extends also to the identification and definition of different features of motion, thereby not only aiming at their implementation but also at a rigorous scientific definition and benchmarking.
Fig. 1

The AMARSi project strives to advance science and technology in the three main directions of mechanics, architecture and learning. The progress in rich motor skills applied to robotics derives from the combination of these orthogonal fields with the bio-inspired paradigms of reservoir computing, adaptive modules and morphological computation

These challenging research objectives are pursued by a consortium of twelve international research laboratories from the fields of bio-robotics, robot engineering, compliant mechanics, morphological computing, human motor research & bio-mechanics, theoretical biology, machine learning, neural networks & reservoir computing. The challenges of merging these different areas are tackled by an intense collaboration programme encompassing seven work packages shared across all research laboratories.

2 AMARSi Towards its Objectives

The AMARSi project started in March 2010 and is currently progressing towards its objectives in three main fields: biology, mechanics and algorithms.

The research on human primitives is contributing to the study of human motion [1, 14], from skills in new born babies [8], to gate transition [15, 37] and recovery [34], crawling [26], anatomy and posture [42], muscle synergies [6], catching [3], motor neuron oscillations [2], spatio-temporal tuning [4] and control behaviors [5]. These studies cast new light on our understanding of how humans learn new motor skills and eventually perform complex and accurate movements. These notions are used in the design of robotics platforms and control algorithms.

The development of compliant actuators for AMARSi has led to novel features in humanoid robots [38, 40] that can now cope with accidental impacts [10], perform fast bipedal walk [23, 24] featuring a range of human-like properties [25, 28]. The COMAN robot, Fig. 2A is currently at an advanced stage of development. Compliant legs have also been adopted in a novel design of a quadruped robot [39], Fig. 2B.
Fig. 2

(A) Legs of the AMARSi COmpliant huMANoid (COMAN). Thanks to cutting-edge compliant mechanics, novel control strategies and learning, this robot can walk, reach and perform various tasks with the smoothness and efficiency typical of humans. See the latest demonstrations on the AMARSi YouTube channel. (B) A prototype of the Oncilla: the quadruped compliant AMARSi robot. This robot can walk softly and smoothly across a variety of surfaces and with different paces in a fashion that resembles our domestic cats and dogs

Morphological computation allows robots to perform movements naturally and efficiently. Progress in AMARSi contributes to morphological computation both theoretically [12, 33] and in robotic design [7, 13], in particular with compliant actuators [27, 36] and evolutionary-designed body parts [16].

While biological notions and hardware components are rapidly developing, the control architectures and learning algorithms are faced with increasingly complex tasks. In AMARSi the application of learning to robotics has further advanced, in particular imitation learning [11, 19], learning of nonlinear systems [17, 18] and learning of motion dynamics for catching [20].

Novel learning methods developed in AMARSi focus in particular on bio-inspired neural networks and reservoir computing. Not only are neural networks biological plausible control structures, their use also helps in understanding important aspects of neural control architectures, how movements are generated, and how learning and plasticity occur. Advances have been achieved in learning the inverse kinematic with reservoir neural computing [32], learning visuo-motor coordination [9], devising control strategies with minimal energy control [22], continuous on line adaptation [21] and human-machine interactive learning [30, 41] as shown in Fig. 3. Networks of spiking neurons have been used to implement probabilistic models [31]. Changing scenarios and behavioral policies have been investigated in [29].
Fig. 3

The AMARSi compliant humanoid robot during a session of kinesthetic teaching. The new compliant mechanics and learning algorithms allow the robot to learn movements from interaction while being guided by a human teacher

The AMARSi partners report ongoing developments currently submitted for publication in all work-packages, Human Motor Primitives, Compliant Systems, Morphological Computation, Adaptive Modules, Learning, Architectures and Software. Public deliverables, open source software, publications and other support material like images and videos are constantly updated on the project website http://amarsi-project.eu.

3 Conclusion

Rich motor skills promise to change the role of robots in our society. Progress in AMARSi shows that robots are becoming less clumsy and are beginning to show increasingly more accurate, more flexible and richer motor behaviors. Such an extended range of behaviors and skills allows robots to perform increasingly complex tasks in diverse human environments like homes and offices. More importantly, smooth and efficient motion under a variety of conditions gift robots with unprecedented human and animal-like resemblance. The contribution of this research project to natural, interactive and safe movements make robots ready to blend into the everyday routines of human society.

References

  1. 1.
    Cappellini G, Ivanenko YP, Dominici N, Poppele RE, Lacquaniti F (2010) Migration of motor pool activity in the spinal cord reflects body mechanics in human locomotion. J Neurophysiol 6:3064–3073 CrossRefGoogle Scholar
  2. 2.
    Chiovetto E (2011) The motor system plays the violin: a musical metaphor inferred from the oscillatory activity of the alpha-motoneuron pools during locomotion. J Neurophysiol Google Scholar
  3. 3.
    Cesqui B, d’Avella A, Portone A, Lacquaniti F (2012) Catching a ball at the right time and place: individual factors matter. PLoS One Google Scholar
  4. 4.
    Christensen A, Ilg W, Giese MA (2011) Spatiotemporal tuning of the facilitation of biological motion perception by concurrent motor execution. J Neurosci 31(9):3493–3499 CrossRefGoogle Scholar
  5. 5.
    d’Avella A, Cesqui B, Portone A, Lacquaniti F (2011) A new ball launching system with controlled flight parameters for catching experiments. J Neurosci Methods 196(2):264–275 CrossRefGoogle Scholar
  6. 6.
    d’Avella A, Portone A, Lacquaniti F (2011) Superposition and modulation of muscle synergies for reaching in response to a change in target location. J Neurophysiol Google Scholar
  7. 7.
    Dermitzakis K, Carbajal J, Marden J (2011) Scaling laws in robotics. In: The European future technologies conference and exhibition Google Scholar
  8. 8.
    Dominici N, Ivanenko Y, Cappellini G, d’Avella A., V., M., Chicchese M., Fabiano A., Sile T., Di Paolo A., Giannini C., Poppele R., Lacquaniti F. (2011) Locomotor primitives in newborn babies and their development. Science 334 Google Scholar
  9. 9.
    Freire A, Lemme A, Steil JJ, Barreto G (2012) Learning visuo-motor coordination for pointing without depth calculation. In: European symposium on artificial neural networks, computational intelligence and machine learning Google Scholar
  10. 10.
    Gan DM, Tsagarakis NG, Dai JS, Caldwell DG (2011) Joint stiffness tuning for compliant robots: protecting the robot under accidental impacts. In: 13th world congress in mechanism and machine science Google Scholar
  11. 11.
    Grollman DH, Billard A (2011) Donut as I do: learning from failed demonstrations. In: International conference on robotics and automation, Shanghai Google Scholar
  12. 12.
    Hauser H, Ijspeert AJ, Füchslin RM, Pfeifer R, Maass W (2011) Towards a theoretical foundation for morphological computation with compliant bodies. Biol Cybern 105(5–6):355–370 MATHCrossRefGoogle Scholar
  13. 13.
    Hauser H, Neumann G, Ijspeert AJ, Maass W (2011) Biologically inspired kinematic synergies enable linear balance control of a humanoid robot. Biol Cybern 104:235–249 MathSciNetMATHCrossRefGoogle Scholar
  14. 14.
    Ivanenko Y, Dominici N, Daprati E, Nico D, Cappellini G, Lacquaniti F (2010) Locomotor body scheme. Human Movement Science Google Scholar
  15. 15.
    Ivanenko YP, Labini FS, Cappellini G, Macellari V, McIntyre J, Lacquaniti F (2011) Gait transitions in simulated reduced gravity. J Appl Physiol 110(3):781–788 CrossRefGoogle Scholar
  16. 16.
    Jones HB, Soltoggio A, Sendoff B, Yao X (2011) Evolution of neural symmetry and its coupled alignment to body plan morphology. In: Proceedings of the genetic and evolutionary computation conference Google Scholar
  17. 17.
    Khansari-Zadeh S, Billard A (2011) Learning stable nonlinear dynamical systems with Gaussian mixture models. IEEE Trans Robot 27(5):943–957 CrossRefGoogle Scholar
  18. 18.
    Khansari-Zadeh SM, Billard AB (2010) An iterative algorithm to learn stable non-linear dynamical systems with Gaussian mixture models. In: Proceeding of the international conference on robotics and automation, pp 2381–2388 Google Scholar
  19. 19.
    Khansari-Zadeh SM, Billard A (2010) Imitation learning of globally stable non-linear point-to-point robot motions using nonlinear programming. In: Proceeding of the IEEE/RSJ international conference on intelligent robots and systems, pp 2676–2683 Google Scholar
  20. 20.
    Kim S, Gribovskaya E, Billard A (2010) Learning motion dynamics to catch a moving object. In: 10th IEEE-RAS international conference on humanoid robots (humanoids), Nashville, TN, pp 106–111 CrossRefGoogle Scholar
  21. 21.
    Lemme A, Reinhart FR, Steil JJ (2010) Efficient online learning of a non-negative sparse autoencoder. In: European symposium on artificial neural networks, computational intelligence and machine learning Google Scholar
  22. 22.
    Li J, Jaeger H (2011) Minimal energy control of an ESN pattern generator. Technical report 26, Jacobs University Bremen Google Scholar
  23. 23.
    Li Z, Tsagarakis NG, Caldwell DG, Vanderborght B (2010) Trajectory generation of straightened knee walking for humanoid robot iCub. In: International conference control and automation, robotics and vision Google Scholar
  24. 24.
    Li Z, Vanderborght B, Tsagarakis NG, Caldwell DG (2010) Fast bipedal walk using large strides by modulating hip posture and toe-heel motion. In: IEEE international conference on robotics and biomimetics Google Scholar
  25. 25.
    Li Z, Vanderborght B, Tsagarakis NG, Caldwell DG (2010) Human-like walking with straightened knees, toe-off and heel-strike for the humanoid robot iCub. In: UKACC international conference on control Google Scholar
  26. 26.
    Maclellan MJ, Ivanenko YP, Cappellini G, Sylos LF, Lacquaniti F (2011) Features of hand-foot crawling behavior in human adults. J Neurophysiol Google Scholar
  27. 27.
    Martinez Salazar HR, Carbajal JP (2011) Including the passive dynamics of a compliant leg in the gait control. In: IEEE/RSJ international conference on intelligent robots and systems, San Francisco, CA, USA Google Scholar
  28. 28.
    Moro F, Tsagarakis N, Caldwell D (2011) A human-like walking for the compliant humanoid COMAN based on com trajectory reconstruction from kinematic motion primitives. In: 11th IEEE-RAS international conference on humanoid robots, Bled, Slovenia Google Scholar
  29. 29.
    Neumann G (2011) Variational inference for policy search in changing situations. In: Getoor L, Scheffer T (eds) Proceedings of the 28th international conference on machine learning. ACM, New York, pp 817–824 Google Scholar
  30. 30.
    Nordmann A, Emmerich C, Ruether S, Lemme A, Wrede S, Steil J (2012) Teaching nullspace constraints in physical human-robot interaction using reservoir computing. In: International conference on automation and robotics Google Scholar
  31. 31.
    Pecevski D, Buesing L, Maass W (2011) Probabilistic inference in general graphical models through sampling in stochastic networks of spiking neurons. PLoS Comput Biol 7(12):e1002, 294 MathSciNetCrossRefGoogle Scholar
  32. 32.
    Reinhart RF, Steil JJ (2011) Neural learning and dynamical selection of redundant solutions for inverse kinematic control. In: IEEE-RAS international conference on humanoid robots (humanoids), pp 564–569 Google Scholar
  33. 33.
    Rueckert E, Neumann G (2011) A study of morphological computation by using probabilistic inference for motor planning. In: 2nd international conference on morphological computation, Venice, Italy, pp 51–53 Google Scholar
  34. 34.
    Solopova I, Tihonova D, Grishin A, Ivanenko Y (2011) Assisted leg displacements and progressive loading by a tilt table combined with FES promote gait recovery in acute stroke. NeuroRehabilitation Google Scholar
  35. 35.
    Soltoggio A, Stanley KO (2012) From modulated hebbian plasticity to simple behavior learning through noise and weight saturation. Neural Network Journal (under review) Google Scholar
  36. 36.
    Sumioka H, Hauser H, Pfeifer R (2011) Computation with mechanically coupled springs for compliant robots. In: IEEE/RSJ international conference on intelligent robots and systems. IEEE Press, New York Google Scholar
  37. 37.
    Sylos LF, Ivanenko Y, Cappellini G, Gravano S, Lacquaniti F (2011) Smooth changes in the EMG patterns during gait transitions under body weight unloading. J Neurophysiol Google Scholar
  38. 38.
    Tsagarakis N, Zhiin L, Saglia J, Caldwell D (2011) The design of the lower body of the compliant humanoid robot cCub. In: International conference on robotics and automation, Shanghai Google Scholar
  39. 39.
    Sproewitz A, KuechlerL, Tuleu A, Ajallooeian M, D’Haene M, Moeckel R, Ijspeert AJ (2011) Oncilla robot—a light-weight bio-inspired quadruped robot for fast locomotion in rough terrain. In: Symposium on adaptive motion of animals and machines Google Scholar
  40. 40.
    Ugurlu B, Tsagarakis N, Spyrakos-Papastavridis E, Caldwell DG (2011) Compliant joint modification and real-time dynamic walking implementation on bipedal robot cCub. In: IEEE international conference on mechatronics Google Scholar
  41. 41.
    Wrede S, Johannfunke M, Lemme A, Nordmann A, Rüther S, Weirich A, Steil JJ (2010) Interactive learning of inverse kinematics with nullspace constraints using recurrent neural networks. In: 20. Workshop on computational intelligence. Fachausschuss Computational Intelligence der VDI/VDE-Gesellschaft Mess- und Automatisierungstechnik, Dortmund Google Scholar
  42. 42.
    Wright W, Ivanenko Y, Gurfinkel V (2011) Foot anatomy specialization for postural sensation and control. J Neurophysiol Google Scholar

Copyright information

© Springer-Verlag 2012

Authors and Affiliations

  1. 1.CoR-LabBielefeld UniversitätBielefeldGermany

Personalised recommendations