Masaaki Mochimaru, AIST:
Introduction of RIKEN-AIST Joint Project

A new RIKEN-AIST joint project entitled "Understanding the mechanism of gene expression through motion activities, and applications to health service for aging societies using VR and robotics" was started in 2017. Over 20 projects were proposed to this new grant, eight proposals were accepted as feasibility studies. After half-year feasibility studies, two projects (including our project) were selected as the full-scale research. The final goal of this project is to support healthy aging life with body remodeling by gene expression through effective daily activities. Especially, we try to support healthy aging life by using interactive systems based on VR and robotics. Effective human activities could be induced by the interactive system. To realize the goal, the project contains the following research topics: (1) understanding the mechanism of body remodeling by gene expression patterns for rodents and macaques; (2) neuro-musculo-skeletal modeling and motion simulation for rodents, macaques and humans; (3) developing interactive systems based on VR and robotics.

Yoshihiko Nakamura, Yosuke Ikegami, Emiko Uchiyama, University of Tokyo:

Performance Analysis of Athletes Using Neuromusculoskeletal 
Human Model

The performance of athletes surprises us by its extremity in power, speed, and accuracy. It is scientific challenges to explain why it is possible from the points of view of mechanics, physiology, neurology, and psychology. We will be able to  gain invaluable knowledge from the challenges on the life of humans. This talk will focus on our recent experience when we had an opportunity to capture and analyze the performance of a top athlete.

Rüdiger Dillmann, KIT / FZI Karlsruhe:

Using Spiking Neural Networks in Closed-Loop Sensory-Motor Robotics Experiments

The execution of sensor guided robot motion primitives and interaction with the environment is still in an early developmental stage compared to highly adaptive and sensor-motor control observed in nature. Humans and animals are capable of adapting their motions and activities successfully to environmental variances, performing meaningful reactions to unforeseen events instantaneously. Flexible and adaptive sensor motor controls, as observed in vertebrates by the interplay of the brain and the spinal cord, is the purpose of our research based on multimodal motion activation using a network of spiking neurons. The functionality of exteroceptive feedback like visual stimuli and motor control is assumed to be represented in higher brain areas.  Reflexes, repetitive low level motion and low level motion primitives are assumed to be represented in the spinal cord. Within SP10 in the HBP a neurorobotics platform has been designed which allows to study closed loop control concepts especially how to generate and to learn them. With various sensory-robotics experiments supported by functional spiking networks basic training, activation stimuli and learning mechanisms of the visual neural layer and the motion activation stimuli are to be studied. Real robots with multi-finger hands as well as spiking DVS cameras are used and connected with the different spiking neural layers.

Silvia Tolu, Technical University of Denmark:

Using robots for understanding the cerebellar role in sensorimotor control

The present research aims to help achieving a better understanding of the brain's interaction with the body and its surroundings, so that the brain is not only understood as a closed, abstract entity. We will describe a modular approach for connecting cerebellar-like models and controling different robotic bodies. The approach is exemplified by a bio-mimetic modular control architecture based on the cerebellar learning concept, machine learning optimisation, and on traditional adaptive feedback control. Here, both robotics and brain science can benefit. The first thanks to a real-time control system that is capable of learning how to perform physical tasks and of adapting to changing conditions, and the second by providing new insights into the the cerebellar involvement in processing the sensory input for motor control tasks. The biologically plausible control system based on internal models and feedback-error learning concepts was tested with different robotic platforms under both manipulation, visual, and locomotion tasks. This will be the state-of the-art for more complex bio-inspired control architectures for neurorobots towards emulating the central nervous system functioning in motor control and learning.

Hidekazu Kaneko, AIST:

Effects of motion-inducing devices on task learning in rats

－Facilitating remodeling in the central nervous system－

Robotic rehabilitation is a research field in neurorobotics that aims to facilitate plasticity of real neural networks by inducing motor sensations. To investigate how motor sensation facilitates learning, we used a sensory–motor association task to determine whether the sensation induced by a motion-inducing device contributed to performance improvement in rats. The rats were trained to respond to a tactile stimulus (air puff) by releasing a lever pressed by the stimulated (compatible condition) or non-stimulated (incompatible condition) forepaw. When error rates fell below 15%, the compatibility condition was changed (reversal learning). An error trial was followed by a lever activation trial in which a lever on the correct- or incorrect-response side was automatically elevated at a preset time of 120, 220, 320, or 420 ms after tactile stimulation. This lever activation induced forepaw movement similar to that observed in a voluntary lever-release response, and also resulted in body movement that occasionally caused elevation of the other forepaw. The latter body movement may have produced a sensation similar to that of voluntary lever release by the forepaw on the non-activated lever. The performance improvement rate was increased by the lever activation procedure on the incorrect-response side (i.e., with the non-activated lever on the correct-response side). Furthermore, the performance improvement rate changed depending on the timing of lever activation: facilitative effects were largest with lever activation on the incorrect-response side, at 320 ms after tactile stimulation, whereas hindering effects were largest for lever activation on the correct-response side, at 220 ms after tactile stimulation. These findings suggest that motion-inducing devices, which provide tactile and proprioceptive stimulation, affect sensory–motor associative learning in a time-dependent manner. We conclude that learning can be facilitated by inducing a motor sensation similar to that resulting from the correct voluntary response, but not by inducing motion. The application of this experimental paradigm to rats with stroke-induced hemiplegia should produce useful knowledge for developing neurorobotic rehabilitation techniques.

Noriyuki Higo, AIST:

Training-induced recovery of hand movements following a focal brain lesion in macaque monkeys

We investigated the effects of post-lesion training on motor recovery after a focal lesion of the primary motor cortex (M1). Behavioral analyses indicated that the recovery of dexterous movements can be promoted by intensive use of the affected hand in post-lesion training. The brain imaging analysis revealed overactivity of the ventral premotor cortex during the recovery period. The causal role of the increased activity in motor recovery was confirmed by means of pharmacological inactivation by muscimol during each recovery period. We also confirmed the plastic changes of neurons in the corresponding areas during the functional recovery using histochemical analysis of a plasticity-related protein. The present results will contribute to develop novel intervention strategies, such as robot-assisted rehabilitation, to restore brain function after brain damage. 

Oleg Gusev, RIKEN:

Tissue remodeling as a biomarker:  promoter-level atlas of diversity of gene expression in the mammalian skeletal muscles

Skeletal muscle plays an important role in many vital processes and is maintained by numerous pathways regulating protein synthesis and degradation. Loss of skeletal muscle mass and strength induces by the various conditions, including muscle disuse resulted in bed rest or unloading, age-related muscle weakness (sarcopenia), cachexia, multiple muscular dystrophies, myopathies, denervation and other malfunctions. Noticeably, muscle atrophy can affect specific fiber types and different muscles have the differential susceptibility to muscle wasting. However, the factors underlying these differences remain to be elucidated.  In the frame of international consortium, we are aiming to reconstruct transcriptional network in variety of muscles in normal and atrophic conditions using Cap Analysis of Gene Expression (CAGE) and Small RNA expression analysis. For the analysis, we are selecting 20-30 target muscles according to several criteria. Firstly, we are considering muscles which is frequently affected or not affected at all by disuse, sarcopenia, cachexia or muscle diseases. The second criterion is an availability of skeletal muscle to a biopsy. Since we are going to elicit age-dependent muscle atrophic changes, two or more age groups will be examined. Besides, male and female comparison will be carried out. It is assumed that as the research objects macaque, subset of human muscles and primary cell lines from patients with dystrophy will be used. As a result, the atlas of promoters and enhancers of RNA transcription in human and animal muscles in normal and pathological conditions will be created. The determination of the transcriptional factors and other genomic regulatory elements involved to differential response of muscle types may lead to a better understanding of the pathogenesis of the various muscle wasting conditions. Finally, integration of gene expression data with rehabilitation procedures will  provide new approach for monitoring of the successful recovery.

Gabriel Urbain, Ghent University  (for Joni Dambre) :

The NRP as a platform for transferable control of a compliant quadruped robot

Although the progress in consumer robotics has been enormous, it remains difficult to combine safety and low cost with anything approaching the adaptivity and dexterity that is desirable for operation amongst humans.

One of the reasons for this is the fact that safe robots, e.g., robots with passive compliance are difficult to control with traditional techniques based on inverse kinematics models. The same is true for robots that are fabricated with low-cost elements, because their sensors and actuators as well as their actual kinematics are less predictable. Due the a large number of small variations such as frictions, spring constants or variability in exact screw positions, relatively large accumulated variations between individual robots may occur. It is exactly for this type of robots that neural control approaches may provide the largest advantages. In order to take the variability into account, at least part of this learning will have to be done on the physical robot. However, in order to be efficient, the actual robot operation time for learning should be minimized. This is why we investigate how to optimally use a closed loop simulation environment like the neurorobotics platform to optimise neural control architectures and provide robust pre-training of a compliant robot platform. As a first test case for this work, we use gait control of the low-cost Tigrillo quadruped platform. We present first results of spiking gait control that is transferable from simulation to the physical robot.

Satoshi Oota, RIKEN:

Of mice and men: a rodent musculoskeletal modeling to bridge genetics and neurorobotics

Our final goal is to develop an endoskeletal robot suit (StillSuit^TM) that ‘remodels’ human mind and body through cognitive and physical interventions. For that purpose, StillSuit^TM needs to know the neuro-musculoskeletal system of an wearer (user) to be remodeled. The remodeling alters the neuro-musculoskeletal system at molecular and macroscopic levels: e.g., gene expression patterns and consequent physiological states. For the detailed analyses on such responses to the interventions, we often have to conduct invasive experiments on subjects. It is virtually impossible to apply a destructive approach on healthy human subjects. Since rodents, including rats and mice, are ideal model organisms to mimic human traits, biologists often analyze them instead. However, there is a huge evolutionary divergence between rodents and human. To translate the animal data to the human traits, we need a powerful yet feasible framework to bridge the species. We are developing multi-scale animal musculoskeletal models and a retargeting theory (described below) for the multi-modal interspecies translation: e.g., from mice to men and from gene expressions to motor functions. Our framework makes it possible to integrate the multi-modal data, through which we study on neurorobotics in terms of the evolutionarily elaborated biological system.

Ko Ayusawa, AIST:

Interspecies motion retargeting for analyzing human body dynamics

The technology of motion analysis can provide various information and insights about the dexterous human movement. In recent days, the detailed human musculoskeletal models have been developed, however, the analysis with such sophisticated human models suffers from various unknown parameters in the computational models. To solve this issue, we aim to develop a methodology of transferring the mechanical knowledge obtained from the experiments of humanoid robots or laboratory animals to humans, by utilizing the technology called motion retargeting.  We introduce two pilot studies: the bi-directional motion retargeting between humans and humanoid robots and the systematic method to find homologous body postures between humans and mice. The works are expected to provide quantitative evaluation framework in order to compare the physical quantities of humans to those of robots and other animals.

Egidio Falotico, Scuola Superiore Sant'Anna:

Towards embedding data-driven brain models in robots

Embedding brain-inspired solutions in real or simulated robotic platforms is a huge challenge that aims to foster our understanding of the human brain.

In the framework of the Human Brain Project (HBP), a collaboration between neuroscientists and robotic engineers has been set up with a twofold goal: 1) merge cutting-edge activities at the experimental and simulation level; 2) build a new conception of closed-loop neuroscience, where experiments drive simulations and simulations guide experimental design. In this new conceptual framework, experiments are built and validated on theoretical models and virtual platforms and vice versa. In addition, the tight collaboration between experimental neuroscientist and model developers is a unique opportunity for a cultural paradigm shift, where the experimental paradigm is better constructed based on theoretical predictions. On the simulation side, neuroscientists created and simulated the first full scale (75 million neurons) scaffold model of a whole mouse brain. This brain model is based on whole-brain imaging data from the Allen Brain Institute as well as the Blue Brain Project.

To make these goals tangible, in the Co-Design project 1 (CDP1) of the HBP, an exemplar experiment based on a paradigm of motor learning, in physiological and pathological settings, respectively has been set up. Activity from the mouse and the mouse brain is recorded before, during and after learning. In detail, the first part is focused on learning of a forelimb pulling task in healthy animals. In the second part of the experiment, a phototrombotic stroke is induced, leading to a loss in motor function. Then, the re-training of the motor task is studied. This part of the experiment is a paradigm for motor rehabilitation after stroke.

In this talk, we will present the current state of the CDP1 experiment, replicated in a simulated robotic environment, the Neurorobotics Platform . This experiment is simulated from the single-neuron activity in the brain and spinal cord to the mouse musculo-skeletal system and to the lab environment. Being able to perform identical experiments in the animal and in simulation, allows us to determine which data or experiment will yield the best improvement of the model and therefore of our understanding.

Akihiko Murai, AIST:

Our final goal is to develop an endoskeletal robot suit (StillSuit^TM) that ‘remodels’ human mind and body through cognitive and physical interventions. For that purpose, StillSuit^TM need to know the neuro-musculoskeletal system of an wearer (user) to be remodeled. The remodeling alters the neuro-musculoskeletal system at molecular and macroscopic levels: e.g., gene expression patterns and consequent physiological states. For the detailed analyses on such responses to the interventions, we often have to conduct invasive experiments on subjects. It is virtually impossible to apply a destructive approach on healthy human subjects. Since rodents, including rats and mice, are ideal model organisms to mimic human traits, biologists often analyze them as model organisms. However, there is a huge evolutionary divergence between rodents and human. To translate the animal data to human traits, we need a powerful yet feasible framework to bridge the species. We are developing multi-scale animal musculoskeletal models and the retargeting theory (described below) for the multi-modal interspecies translation: e.g., from mice to men and from gene expressions to motor functions. Our framework makes it possible to bridge the multi-modal rodents and human data, through which we study on neurorobotics in terms of the evolutionarily elaborated biological system.

Jun Igarashi, RIKEN:

Introduction of large-scale neural network simulations in the project for a next-generation supercomputer in Japan

Large-scale simulations of realistic neural networks have been actively conducted thanks to increases in computing performance according to Moor’s law and increase in available physiological data due to improvements in observation technology in neuroscience. Next-generation Japanese supercomputer, “Post-K” is planned to appear during 2021-2022. The performance of the supercomputer is going to be 10-100 times larger than that of the current supercomputer K that has 11 petaFLOPS of theoretical performance. We have worked on the project of a large-scale simulation of a whole-brain model consisting of cortex, cerebellum, and basal ganglia using the “Post-K” supercomputer for understanding information processing of movement behaviors and thinking in the mammalian brain. In this talk, we will introduce the project and the research activity focusing on the study of parallelization of layered cortical sheet model executed on K computer.

Shingo Shimoda, RIKEN:

Symbiotic interaction with robots for supporting human behaviors

Efficient interactions between human and robots are long-standing issue. Recent advances on this problem is the deepr understandings of neural plasticity during the interactions with srounding enviornments. It is getting clear that there exists the appropriate interaction that our neural plasticity works efficiently to adapt the behaviors to the new environment. If incorrect mechanical supports are provided, for example, our original control system fights with the external support resulting unexpected mismaching. To overcome this problem, we proposed the biomimetic learning algorithm called tacit learning to control the robot to adapt to humans’ behaviors. We introduce the control of forearm prosthesis and walking support exoskeleton robot according to user’s motion intentions through robot/body interactions in this workshop.

Eiichi Yoshida, AIST:

Human-Centered Product Design and Evaluation through Integrated Motion Analysis using Digital Human and Humanoid 

We present an integrated approach to motion analysis for humanoid robotics and human simulation that is useful for product design and evaluation. The first axis is development of a method for humanoid robot control that can reproduce various human behaviors to use a humanoid robot as an evaluator of products such as assistive devices. This allows estimating its mechanical supportive effects in a quantitative manner, which is difficult with human measurement. We also introduce applications of this research to standardization of wearable lumbar-support assistive devices. Another main research direction is to develop a system for human-centered product design through understanding humans' motion principles by using a digital human that can model its shape, musculo-skeletal structure and motions, as well as interactions with devices. We will show some applications of product design and evaluation based on this research.

Hirohisa Hirukawa, AIST:

Elderly-Care Robots for Normalization

Japan has been transferring to aged society rapidly, and the management of elderly-care is going to a big challenge before 2025.  The Japanese government has a national project to develop nursing-care robots from 2013 to 2018, and fifteen commercial robots have been developed.  The robots include powered suits for care givers, transfer assist among bed, wheelchair and toilet, walking assist for outdoor and indoor, smart toilet and bath lifts.  The objectives of the robots are to reduce the burden of care givers and to enhance the autonomy of elderly persons.  The project will focus the enhancement of the autonomy, and soft powered suits will be developed to this end.  This talk introduces the project with the emphasis on the enhancement of the autonomy towards the normalization of the life of elderly-person.