Started development of autonomous task demo using DOBOT's collaborative robot arm "MG400"

■ Overview

EmplifAI, Inc. (Headquarters: Bunkyo-ku, Tokyo, Representative Director: Hiroyuki Osone, hereinafter referred to as "EmplifAI") has begun developing a control and learning stack (hereinafter referred to as "MG400 LeRobot Stack") using the open source robotics platform "LeRobot" for DOBOT's collaborative robot arm "MG400".

This project is being carried out using the ``DOBOT MG400'' on loan from Afrel Co., Ltd., an authorized domestic distributor of DOBOT products.

Operation video

In this development, we aim to use MG400 as a physical AI platform that can ``think and act like a worker in the field'' by combining ``imitation learning'' and reinforcement learning, which learn robot behavior from human operation data.

In the future, we will continue to develop the product in stages, with a view to displaying it at actual demonstrations and exhibitions.

■ Development background and purpose

In manufacturing, logistics, and research and development fields, labor shortages and skills inheritance issues have become apparent, and there is a need to break away from the situation in which only people who can write programs can run robots.

On the other hand, although collaborative robot arms are becoming more popular around the world, the software stacks are separated by manufacturer, making it difficult to link them with AI and imitation learning.

While providing support for the development of physical AI for humanoids and collaborative robots, EmplifAI strongly felt the need for a "robot learning stack compatible with multiple manufacturers" that can consistently handle everything from data collection to learning and deployment using a similar workflow for robots from any manufacturer.

As a first step in this project, we will work on building a development platform for DOBOT's MG400, which is compact and easy to use, in combination with LeRobot, an open source robotics library.

■ Initiatives for “MG400 LeRobot Stack”

In this project, we are building a development stack with the following functional groups.

Teleoperated data collection pipeline

Operators intuitively operate the MG400, and its trajectory, joint angles, camera images, etc. are automatically logged.

Policy learning using imitation learning and reinforcement learning

It converts the collected data into a LeRobot compatible format and learns autonomous movement policies for tasks such as object grasping, alignment, and simple assembly.

Deploying the trained model to the actual MG400 machine

We have established the infrastructure to bridge the learned policy to the MG400 control interface and perform operational verification and tuning on the robot.

Architecture with multi-manufacturer deployment in mind

Based on the current MG400 compatibility, we have standardized the data format and API design so that it can be horizontally extended to other companies' arms and humanoids in the future.

Through this, we will realize the use of physical AI from the perspective of the field.

■ About EmplifAI Co., Ltd.

Company name: EmplifAI Co., Ltd.

Address: Sun Court Yushima 1F, 3-21-5 Yushima, Bunkyo-ku, Tokyo 113-0034

Representative: Representative Director Hiroyuki Osone

Business content: Physical AI/robotics development support, AI control algorithm development, robot introduction/PoC support, humanoid robot rental, event exhibition, etc.

URL：https://emplif.ai/

■ Contact information regarding this matter

E-mail：support@emplif.ai