| 研究実績の概要 |
We described a methodology to automatically design chemical controllers to drive swarms of self-propelled biochemical-micro-robots to self-organize through local interactions and self-assemble into specific shapes. This task is challenging due to very high-dimensional sparse and deceptive search spaces, non-linear multi-scale dynamics and high experimental variability. We tackled the problems in simulations by using quality-diversity algorithms. In particular, we used the MAP-Elites algorithm to find a diverse set of high-performing chemical reaction network controllers to drive the robot swarm. These results were presented at two international conferences: SSCI 2019 and SWARM 2019.
Moreover, we assessed the performances of MAP-Elites to find good controllers for other robotic problems. These results were presented at the GECCO 2019 and GECCO 2020 international conferences and published in the associated proceedings.
In collaboration with our partners at The University of Tokyo, we investigated how swarms of biochemical-micro-robots can be used for a different task, namely to sense and identify collectively the geometrical shape of their environment. The swarm is put into reactors of given shapes (e.g., a '1'-like shape, or a '3'-like shape), and estimates the geometrical features of the reactors through reaction-diffusion and Turing pattern formation. This approach make use of a large number of chemical species that interact in way inspired by convolutional neural networks, with edge detector, pooling, and linear classifier dynamics.
|
| 今後の研究の推進方策 |
This year, we plan to first consolidate our previous results where DNA strands robots controllers were optimized by quality-diversity algorithms to self-assemble into target shapes. We will then further improve our approach by using Machine Learning techniques (e.g., graph autoencoder) to reduce the complexity of the search space of the problem. This would allow us to scale our methodology to more complex target shapes, and simulate more complex robotic behaviors. Namely, we will extend the range of target shapes to more complex shapes composed of several sub-components, possibly with temporal dynamics (moving shapes).
All previous works were performed in simulation, and we plan to validate them experimentally this year on two types of swarm robotic platforms, composed either of micro-beads coated by DNA (Biochemical-micro-robots), or of centimeter-sized mechanical robots. The experiments with the Biochemical-micro-robots will be realized in collaboration with our colleagues at The University of Tokyo and Tohoku University. The experiments with the mechanical robots will be performed with our partners in France at the Sorbonne University.
We plan to publish all research results in international journals and present them at international conferences.
|
