Yoshihiko Nakamura, Yosuke Ikegami, Kazunari Takeichi, Akihiko Murai |
The efficient mathematical algorithms for kinematics, dynamics, and optimization have been developed in robotics and applied to realtime-critical problems of robot control. They have contributed to biomechanical analysis of human movements. At the University of Tokyo our group with Dr. Katsu Yamane developed sDIMS, a computational library including an original parallel algorithm for forward dynamics computation and an original LCP solver. It has been applied to biomechanical analysis using the human musculoskeletal model with over 150 degrees of freedom and nearly 1000 wire muscle model. The inverse dynamics computation and optimization algorithms execute and visualize the muscle activities of the whole body in realtime at the rate of 30 frames per second. The technology is called “Magic Mirror.” The spinal reflex model was first discussed with the biomechanical musculoskeletal model in the Ph.D dissertation of Akihiko Murai in 2009. The neuro musculoskeletal computation in supercomputing frame work started in 2011 and has shown some remarkable results, which include the finite element computation of the whole-body musculoskeletal system and the motor neuron pools in the spine and the muscle-spindle sensory neuron pools (Kazunari Takichi, Shoken Hirasawa, Akihiko Murai, Ko Ayusawa). The neuron pools are modeled based on the literature information and the estimate based on its interpolation/exterpolation using PyNEST for massively parallel computation. This model will be combined with the brain model of the Okinawa Institute of Science and Technology for simulating syndrome of the Parkinsonians. Some preliminary results will be shown and discussed. |
Rüdiger Dillmann |
This talk is discussing the requirements, the potential benefit and the expectations to the simulation environment for neuro-robotics being controlled by neural brain-like networks. The spectrum of robots ranges from walking machines, robots with artificial muscles, anthropomorphic robots or dexterous manipulation. The architecture and the components of the Neuro-Robotics Platform developed in the Human Brain project will be presented and discussed. The robotic systems and the ongoing research on anticipation based on neural control architectures are presented. The six-legged walking robot LAURON is a research platform with a focus on machine learning, especially classical neural networks. The early systems have been used to create swing stance leg cycles and the coordination of walking patterns as well as stabilization reflexes. The current generation, LAURON V, now applies a behavior-based control architecture. Its successive transition to a control architecture based on spiking neural networks is subject to ongoing research focusing the generation of the control and generation of walking patterns considering the perception of the environment. Further test platforms include the anthropomorphic service robot HoLLie that consists of two light-wight arms and dexterous anthropomorphic five-finger hands as well as a mobile KUKA platform with 2 redundant multi fingered lightweight arms. The talk concludes with research related to learning especially learning by demonstration, dexterous manipulation and grasping and perception, together with achieved results (DEXMART, DESIRE, PACO+). This aims to open a discussion on how to design anticipative systems in the context of neural control architectures and its evaluation by the HBP robot simulation framework. |
Yasuo Kuniyoshi |
How does human mind develop? What causes developmental disorders? Development is a continuous bootstrap process of complex interaction between genes, body, nervous system and environment. Therefore investigating the global structure of the process from the beginning is crucial for understanding the fundamental principles of human development. With the advent of “4D” ultrasound imaging and fetal MRI, a burst of data has been accumulating about human fetal development. Also, increasing number of reports suggest that perturbation or abbreviation of fetal development may be relevant to later developmental disorders. Besides, compared to infants/toddlers in extremely complex environment, fetuses may allow more principled way of modeling. We constructed a simulation model of a human fetus [Kuniyoshi&Sangawa, Biol.Cybern. 2006]. It consists of a musculo-skeletal body, uterus, and basic nervous system. It exhibits spontaneous motor development and sensory-motor map organization comparable to human data. Also, by changing the model parameters, we can simulate “abnormal” development [Mori&Kuniyoshi, ICDL2010]. A series of such experiments suggest that sensory-motor experiences in the fetal period may be crucial to the formation of body schema, and that some conditions characteristic of preterm infants lead to degraded body representations [Yamada et al. ICDL-EpiRob2013]. In 2012, we started a large scale project for establishing “Constructive Developmental Science”(http://devsci.isi.imi.i.u-tokyo.ac.jp/), a truly trans-disciplinary research field integrating robotics, medicine, psychology, neuroscience, and “Tojisha-Kenkyu” (first-person view research of developmental disorders). Its main contributions include a new understanding of human development and its disorders, comprehensive diagnostic methodologies, and truly appropriate (for Tojisha) assistive technology. Our first goal of research is to understand how embodied sensory-motor processes relate to social cognition, and to reveal the early causes and their unfolding towards ASD. Our fetal development simulation serves as a platform for integrating the trans-disciplinary data/models and experimenting on the plausible developmental trajectories under alternative conditions. Constructive developmental science Web site. http://devsci.isi.imi.i.u-tokyo.ac.jp/ (2012) Kuniyoshi, Y., Sangawa, S.: Early motor development from partially ordered neural-body dynamics– experiments with a cortico-spinal-musculo-skeletal model. Biological Cybernetics 95(6), 589–605 (2006) [Mori&Kuniyoshi 2010] Mori, H., Kuniyoshi, Y.: A human fetus development simulation: Self-organization of behaviors through tactile sensation. In: IEEE 9th International Conference on Development and Learning (ICDL 2010), pp. 82–97 (2010) [Yamada etal 2013] Yamada, Y., et al.: Impacts of environment, nervous system and movements of preterms on body map development: Fetus simulation with spiking neural network. In: IEEE ICDLEPIROB (2013) |
Stefan Ulbrich |
This talk presents the research roadmap and related prior work of the learning group at FZI in the scope of the Human Brain Project. The main goal is the study and design of an engineered, coherent neural system consisting of well distinguishable sub-components that are neural counterparts to well understood approaches from the field of robotics and classical machine learning. This bottom-up approach intends to fill the gap between applied robotics and neuro-informatics research. On the one hand, this implies a reliably working sensorimotor system that can be monitored. On the other hand, this research fosters the identification of the functionality and processing principles of neural structures in the human brain. In this talk, the Kinematic Bézier Maps (KBM) are presented as an example for body schema learning. They are a class of mathematical models that can represent robot kinematics and dynamics in a linear form that can be thus learned very easily. Because of indications for the presence of parameter transformations in the sensorimotor processing of humans, they are considered a basis for our research. Given the fact, that the learning speed of the KBM reaches the theoretical upper bound, they are the most promising approach that, above all, can be easily expressed in a neural form. From this starting point, cortical maps will be constructed that can represent the whole body, can be combined with visual input, and play a role similar to the mirror neurons in the human brain. With the progress of research, the system will be combined with more components such as the processing of visual stimuli and memory using state-of-the-art techniques (e.g., deep learning) that are to be connected to the proposed sensorimotor system. Altogether, these subsystems will be combined into a greater, coherent system while this neural engineering will also focus on biological plausibility. |
Shu Takagi, Naoto Yamamura, Kazuya Shimizu |
We have been developing a multiscale computational model of the skeletal muscle for the supercomputer “K”. The model is capable of coupling the muscle behavior with the cerebral nervous system, aiming to provide a useful tool for investigating the mechanisms of motor dysfunctions appearing in Parkinson’s disease, and exploring effective therapeutic approaches. Here, a three dimensional finite element (FE) model of the musculo-skeletal system is developed with the spike signal from the spinal cord taken into account. That is, the each muscle fiber receives motor commands from the cerebral nervous system and gives the contraction of fiber. This contraction of each muscle generates a contraction of the entire muscles and gives the joint movement as a result. Furthermore we introduced a muscle contraction model based on microscopic stochastic sarcomere kinetics to the musculo-skeletal FE simulator. The stochastic behaviors of the sarcomere, such as cross-bridge kinetics, are reproduced by conducting Monte Carlo (MC) simulations for the transition state models of the T/T (troponin/tropomyosin) units on the thin filament and the myosin molecules on the thick filament. The simulation results using the above hierarchical integrated mode of skeletal muscle is discussed in the present talk. |
Florian Walter |
Towards understanding the evolution of human bipedalism based on neuro-musculoskeletal simulation” Mastering the challenges of a complex, unpredictable and continuously changing environment is the most basic task of every living creature. At the same time, it is one of the hugest unsolved problems in robotics research. In this context, machine learning has been early recognized as a means of overcoming the limits of classical mathematical models by trying to imitate the cognitive capabilities of humans and animals. Since its early beginnings, machine learning has evolved into an own field of research influenced not only by robotics but also many other disciplines like optimization, statistics, information theory, biology and neuroscience. The presentation gives an overview of this diverse field with a special emphasis on links to robotics and cognitive systems. In particular, it addresses two different but nevertheless coherent perspectives. In the first part of the talk, we recapitulate common machine learning methods and theories based on optimization, statistics, formal logic and artificial neural networks. An important result of this review will be the importance of appropriate representations of the input data for the corresponding learning algorithm, which motivates the discussion of deep learning techniques. The second part of the presentation investigates neurobiological learning techniques from theoretical neuroscience which are based on highly realistic mathematical models of biological neurons. We will both illustrate links of this field to classical machine learning and point out the specific advantages of spiking neural networks competing concepts. Since only realistic input and output data will allow the emergence of realistic brain-like neural dynamics, special emphasis will be put on illustrating the importance of neurorobotics implementations of the discussed learning techniques. |
Naomichi Ogihara |
Fundamental difficulties exist in attempts to clarify the origin and evolution of human habitual bipedal locomotion based solely on morphological analyses of hominin fossils as such fossils are spatiotemporally very scarce. In order to complement our understanding of the evolution of human bipedal locomotion, we have investigated bipedal walking of Japanese macaques as a living analogue for the earliest protohominids who had just started to walk bipedally, as their musculoskeletal structure was presumably not adapted to bipedal walking. However, elucidating the complete mechanisms of bipedal walking based solely on experimental analyses is not trivial. Therefore, we recently employ computational techniques to investigate causal relationships among morphology, kinematics and energetics of bipedal locomotion based on a predictive neruo-musculoskeletal simulation. This would enable evaluation and prediction of changes in kinematics, kinetics, and energetic of locomotion resulting from virtual alterations of the musculoskeletal system, such as deformation of skeletal morphology or modification of muscle length, size, or disposition. Therefore, a forward dynamics simulation of a musculoskeletal model would be an effective way to examine hypotheses and scenarios of the origin and evolution of human bipedalism. |
Joni Dambre |
Motor control systems in the brain humans and mammals are hierarchically organised, with each level controlling increasingly complex motor actions. Each level is controlled by the higher levels and also receives inputs from the senses and from the body itself. Through learning, this hierarchical structure adapts to its body, its sensors and the way these interact with the environment. An even more integrated view is taken in morphological or embodied computation, by stating that the way the body is constructed can drastically simplify motor control and that the body should be considered as an integral part of the control. Ghent University’s Reservoir Lab was named after its original research domain: Reservoir computing. This technique allows to maximally exploit the computational power of recurrent neural networks with minimal training. It was originally proposed in the flavours of echo state networks (analog neurons, H. Jaeger) and liquid state machines (spiking neurons, W. Maass), both of which show similar computational properties and are therefore in principle interchangeable. My talk will give an overview of our work on hierarchical motor control in compliant robots using echo state networks (because their simulation is computationally less intensive. Throughout the years, we have been involved in the design of a range of compliant robots. For these platforms, we have used reservoir computing for creating appropriate tuneable CPGs, for training feedback controllers that maximally exploit the body dynamics (embodied control), and self-organising forward and inverse models. A large part of this work has been performed in the international context of the European AMARSI project. |
Satoshi Oota |
The mouse is an excellent model organism for studying human diseases. Meanwhile, human and mouse diverged more than 90 million years ago and independently accumulated mutations in their genetic information, resulting in phenotypic gaps between them. The gaps, which are sometimes called the extrapolation difficulty from mouse to human, are more serious at macroscopic (whole-body) level rather than microscopic (molecular and cellular) level, due to evolutionary divergence of neuro-biomechanics traits: e.g., neural control and musculoskeletal topography, as well as physical properties. Since we preferentially use model organisms to investigate the whole-body level phenomena, the extrapolation difficulty is a serious and somewhat ironical issue. Considering that the mouse is expected to mimic human traits, this is a grave situation, especially when we are interested in aberrant neuro-motor functions caused by neurological diseases. We are developing a fine-grained neuro-musculoskeletal model of the mouse to bridge the two species at the biomechanics and neural control level. Assuming that the mouse and human share the basal musculoskeletal topology, we defined evolutionarily relevant landmarks between their skeletal systems, and deformed the human skeletal model to the mouse. Next, we mapped a known human muscle topology to the mouse skeletal model by using the deformation function, interpolating attachment sites of muscle architecture units on the skeletal system. To acquire fine topographical information of the mouse musculoskeletal system, we also used a transgenic mouse strain carrying the GFP reporter gene under control of the Scleraxis promoter (Scx-GFP), which is exclusively expressed in the tendons and connective tissues. Our fine-grained mouse neuro-musculoskeletal model can potentially make it possible to extrapolate observed neuro-motor functions from mouse to human at motor command level. Our final purpose is to integrate our fine-grained mouse neuro-musculoskeletal model with the brain model, which will be developed in the human brain project. With the integrated mouse brain-body (IMBB) model, it is possible to coherently perform forward simulation of various mouse behaviors governed by the central nervous system. By bridging the human and mouse neuro-musculoskeletal models, we will interpret the mouse behaviors to human at motor command level. Furthermore, with the IMBB model, it is potentially possible to associate the human neural controls with the actual (observed) motor functions. We should note that the mouse is an excellent genetic tool, by which we can manipulate the organism at genetic level: e.g., it is even possible to control genetically-modified mouse behaviors by using optical stimuli to the nervous system (optogenetics). In our framework, such rich data will be applicable to construct a biologically relevant central nervous system model. Of course, the IMBB model can be a good tool to “annotate” the connectome of the brain especially in terms of neuro-motor functions. Every extant organism on the earth has survived for 4 billion years: the genetic information of animals should harbor secrets of robust and intelligent motor functions. I believe that our project has great potential for cooperation with the human brain project, especially with sub-project “Nerorobotics.” |
Jun Igarashi |
Parkinson's disease (PD) is a degenerative disorder of central nervous system, caused by a loss of dopaminergic neurons in substantia pars compacta. The loss of dopamine causes abnormal synchronized neural oscillation in the basal ganglia in the frequency rage of 8-30 Hz, which is thought to influence to the thalamus and cortex and cause PD motor symptoms, akinesia, bradykinesia, tremor, and rigidity. We have been developing a model of basal ganglia-thalamocortical circuits for understanding developmental mechanism of Parkinson's disease motor symptoms and information processing mechanism of motor behaviors. Physiological experiments by Tachibana and his colleague (2011) suggested that the abnormal neural oscillation may occurs in the interaction between the globus pallidus external segment (GPe) and the subthalamic nucleus (STN) in the basal ganglia. To investigate affects of strength of connections, stochasticity of synapses, and spontaneous firings to the abnormal oscillation, we developed a model of GPe and STN using conductance-based neuron models and stochastic synapse models. When we set concentration of dopamine to PD state in the model, abnormal oscillatory spiking which had peak of power at about ~14 Hz appeared in GPe and STN neurons. The results suggest that firing property of STN neurons and decrease in spontaneous firings of GPe neurons may be critical in the occurrence of the abnormal oscillation. The origin of PD resting tremor has been suggested to be in the brain. However, in terms of frequency range, 8-30 Hz of the before-mentioned abnormal oscillation in the basal ganglia is different from ~4 Hz of PD resting tremor. In addition, two electromyographs (EMGs) recorded from extensor and flexor muscles during PD resting tremor shows anti-phase synchronization of ~4Hz oscillations. It has not been known how PD resting tremor occurs with these phenomenon through the basal ganglia-thalamocortical circuit. In order to investigate these, we developed spiking neural network model of thalamus and motor cortex according to the information of cell density, spatial extent of connections, and layer structures. The model of motor cortex showed selection of neural activities corresponding to the size of column due to lateral inhibition. The model of thalamus showed anti-phase synchronizations of low-threshold Ca 2+ spikes of thalamocortical neurons when thalamic neurons were inhibited assuming PD state. Finally, the combined thalamocortical circuit showed that anti-phase synchronized oscillatory firings of pyramidal-tract neurons of different columns, which was similar to EMG signals appearing in PD resting tremor. These result suggest that the frequency change between the oscillations in the basal ganglia and PD resting tremor, and anti-phase synchronization of EMGs of antagonistic muscles might occur due to activation of low-threshold Ca 2+ spikes in thalamic neurons and lateral inhibition in thalamus and motor cortex. We also tested computational performance of large-scale integrated model of the basal ganglia-thalamocortical circuits on K computer with NEST and MUSIC. The scale of the model achieved full spatial size and neuron number of basal ganglia, thalamus, and motor cortex of the rat, which included more than 3 million of neurons. |
James Knight |
Over recent years, neuroscience has provided many insights into how biological brains process sensory data and provide motor control, but what of the electronic ‘brains’ that would allow these insights to be translated into future neurorobotic systems? It seems natural that neuromorphic hardware platforms could provide the solution. In this talk, I will outline the features of the SpiNNaker system and explain how these features will permit the construction of biologically inspired neurorobotic systems. SpiNNaker is an extensible massively-parallel computer system, and available systems range in size from individual chips with just 18 cores up to the full system with over a million ARM processor cores (the largest machine to date has 100,000 cores) connected using an innovative lightweight packet-switched communications fabric capable of supporting typical biological connectivity patterns in biological real time. Compared to the usual neuromorphic approach - which is implemented using analogue electronics - SpiNNaker is much more flexible, since the simulation kernels running on each processing core are entirely implemented in software. This flexibility, combined with its real-time nature, makes SpiNNaker an ideal platform for neurorobotics: not only can it simulate large-scale neural systems using a wide range of neural and synaptic models, but its processing cores can also be used to interface with a wide variety of sensors and actuators. Of course there's no such thing as a free lunch, and digital neuromorphics are less energy-efficient than their analogue counterparts. My particular interest lies in how we can implement biologically-inspired synaptic plasticity rules on this system; I will explain in outline some of the biological processes involved in synaptic plasticity, and how we can simulate this activity in software. This work has incorporated ideas from computational neuroscience simulators such as NEURON and NEST, in particular our partners in Julich and KTH. |