1. HOME
2. RESEARCH
3. SUMMARY

Integrated control of base construction with multiple robots

This project aims to realize a self-evolving AI system embedded in a versatile modular multi-agent robot system. Deep reinforcement learning has been studied and has yielded results as an AI for generating and controlling robot movements. However, current research achievements primarily focus on single-body robots or the implementation of individual task learning. To apply these advances to reconfigurable modular robots and heterogeneous robot groups, it is necessary to establish methods that enable the Plug and Play (transfer, reuse, reconfiguration) of learned outcomes. Developing hierarchical reinforcement learning is a promising approach. In this research and development task, we will evaluate the AI technologies developed, particularly using assembly tasks.

Implementation of Distributed AI system with Hierarchical Reinforcement Learning

As an achievement of this fiscal year, we focused on a robot system equipped with hand and arm modules. We evaluated the learning performance when the hand robot and arm robot were combined and transformed into different robot configurations. Specifically, we compared the performance of using learning data from the hand robot alone and the arm robot alone with the performance of learning from scratch with the integrated hand-arm robot. As a result, we found that by using the learning data, manipulation tasks could be completed with much fewer trials.

RESEARCH & DEVELOPMENT ITEMS