This time, Macnica exhibited under the theme of "Three Computer Solutions" advocated by NVIDIA.

A Three-Computer Solution: Powering the Next Wave of AI Robots

For more details about these "three computers," please check the website above, but to briefly explain the overview,

① "Learn" (train) the AI model using large amounts of data

② "Test" the robot in a virtual space (simulation)

3. The learned and tested content is then "implemented" (deployed) on the robot itself.

The process is as follows.

By utilizing AI, we believe that we can move away from conventional rule-based robot operations and create robots that can handle a variety of uses (multitasking).

At this exhibition, we will be showcasing some of these solutions.

This exhibition introduced examples of "testing" and "dataset generation" using robot simulation in a virtual space, as well as "dataset expansion," which automatically generates variation videos by replacing objects in the generated dataset video.
In the "Dataset Generation" section, we exhibited a demo of Pick&Place, which recreates robots and work environments in a virtual space and simulates real-world work scenarios. This utilizes the NVIDIA Omniverse™ library and NVIDIA Isaac Sim, an open-source robotics simulation framework. Data that is difficult to collect on-site can be safely collected in large quantities anytime and anywhere in a virtual space.
Furthermore, we used the NVIDIA Cosmos™ world-based model to expand the dataset video by adding various conditions and variations.
Dataset generation/expansion improves the versatility and adaptability of AI models, enabling the development of robust robotics AI that can withstand unexpected situations in real-world environments.

Simulations are performed in both virtual and real spaces to evaluate movement, precision, reaction speed, etc. This is a technical verification that is directly linked to actual robot operation.

Utilizing NVIDIA Isaac Manipulator, a software platform designed to accelerate the development of robotic arms, the system uses robot and obstacle information obtained from camera footage to generate the optimal path between two given poses in just a few seconds, allowing the robot to avoid unknown obstacles.
In this demonstration, the Gripper attached to the end of the arm is also controlled at the same time to perform pick and place operations while avoiding obstacles.
As an example, consider the automation of pick-and-place tasks using a robot on a production line in a factory. However, if the robot encounters an obstacle during the task, it will have to be stopped, resulting in reduced production efficiency.
Using this technology, obstacle avoidance becomes possible without teaching, making it possible to maintain production efficiency.

Using the NVIDIA Isaac GR00T™ N1.5 Vision Language Action (VLA) model, the robot arm SO-ARM101 demonstrates a series of tasks, picking up an object and placing it accurately in a specified location.
Isaac GR00T N is the world's first "open platform model for humanoid robots" developed by NVIDIA. Although it is designed for humanoid robots, it can also be used for robotic arms.

It is noteworthy that this demo was built using data from only about 30 teleoperation operations. GR00T N1.5's versatile inference capabilities and pre-training on large-scale data have made it possible to implement this with such a small dataset.

For this exhibit, the GR00T N1.5 has been finetuned and deployed on NVIDIA® Jetson Thor™, the latest Edge AI platform released in August 2025. Jetson Thor, designed for robotics, supports the robot's performance.

NVIDIA Jetson Thor - NVIDIA -Macnica