Theme- 1 橋本班 Team HASHIMOTO

Development of χlogbot and elucidation of behavioral diagram using sparse modeling

To analyze a behavioral diagram, we must predict the purpose of the behavior. Mathematically, this is an inverse problem of estimating an optimal evaluation function from which the observed behavior is derived as a solution. Behavior must be related to the interaction with the environment and the individual's internal state; however, it is not easy to measure all causes of a specific behavior. Because sparse modeling finds appropriate relationships from less data, it is effective in such cases. The inverse problem, in which the optimization parameters are estimated from the data, can be formulated as a combinational optimization problem. In general, to obtain an exact solution while estimating the evaluation function, it is necessary to perform an exponential number of repetitive calculations for a large number of parameters. On the other hand, quantum annealing calculations using the Ising model can lead to a fast and sparse approximate solution (which expresses the behavior using the minimal number of motion primitives).

This research aims to formulate an estimation problem for behavioral diagrams using sparse modeling and achieve a high-speed solution. As for the χlogbot, we are developing an interactive projection mapping system with Kagami contributors and are working on behavior analysis and real-time intervention. We also work on behavioral modeling with Ando, and will jointly develop a robotic microscope and conduct a precise analysis of stimulus interventions. In collaboration with Team Kawashima and Team Fujii in A02, we will develop basic technology to visualize the structure of behavioral diagrams. In addition, based on robotics, control engineering, and analysis of behavioral diagrams, we will design, prototype, and develop the hardware of the χlogbot that enables on-board analysis and intervention with Team Makino in A02 and Team Takahashi in A01, and develop the χlogbot software in cooperation with Team Maekawa in A02. We will also collaborate with A01 members, Team Yoda, Team Hiryu, and Team Nishimori to obtain real-world data.

• 

A02-1 橋本班 研究者

• 

  supervisor
  Principal Investigator Koichi Hashimoto Koichi Hashimoto

  Graduate School of Information Sciences, Tohoku University Professor

  Research field

  Robotics, Computer Vision, Control Engineering

  Research keywords

  Visual Servo, Point Cloud, Deep Learning, Trajectory Mining, Visual SLAM

  Laboratory

  http://www.ic.is.tohoku.ac.jp/en/koichi/

• 

  Co-Investigator Keiko Gengyo-Ando Keiko Gengyo-Ando

  Graduate School of Dentistry, Tohoku University Associate Professor

  Research field

  Study on neural circuit mechanism of behavior using in vivo imaging of neural activity.

  Research keywords

  Calcium imaging, behavior, neural circuit, C. elegans, molecular neurophysiology

  Laboratory

  http://www.dent.tohoku.ac.jp/english/field/morphology/02/index.html

• 

  Co-Investigator Shingo Kagami Shingo Kagami

  Unprecedented-scale Data Analytics Center, Tohoku University Professor

  Research field

  Sensory information analysis including vision and image processing

  Research keywords

  Robot Vision, High-Speed Vision, Low-Latency Display, Augmented Reality, Projection Mapping

  Laboratory

  https://sites.google.com/view/kagami-lab-tohoku-u/