Combining human teleoperation demonstration learning, large-scale simulated data generation via NVIDIA Omniverse / Isaac Sim, and the NexCOBOT MARS400 Pro edge deployment platform powered by NVIDIA Jetson Thor, to build a complete embodied-AI technology stack spanning data collection, simulation training, and real-world autonomous execution

TAIPEI, August 2026 — NexCOBOT and ITRI jointly unveiled a next-generation robotic AI training and deployment technology, “The Next-Gen AI Controller,” at the recently concluded 2026 TAIROS (held August 19–22). Starting from human teleoperation demonstrations, the technology uses NVIDIA Omniverse and Isaac Sim to scale real-world data into large simulated datasets. The resulting Vision-Language-Action (VLA) model is ultimately deployed on the NexCOBOT MARS400 Pro edge computing platform, powered by NVIDIA Jetson Thor, giving the robot real-time autonomous decision-making and execution capability — bridging the technology gap from data collection to real-world deployment.

In this demonstration, NexCOBOT provides the edge computing and robot control platform powered by NVIDIA Jetson Thor, while ITRI contributes AI learning, model training, and Real-to-Sim-to-Real technology. Together with NVIDIA’s Omniverse, Isaac Sim, DGX, and GR00T AI development and simulation tools, the collaboration presents a complete technology pipeline from data collection and model training through to on-robot deployment.

News Summary

• Complete Real-to-Sim-to-Real loop: Starting from human teleoperation demonstrations (Imitation Learning), precise robot motion ground-truth data is collected as the training foundation
• Large-scale simulated data generation: NVIDIA Omniverse and Isaac Sim scale real demonstration data into a highly randomized virtual training dataset, combined with NVIDIA Cosmos and Mimic/Gen for data generation and augmentation; training data is stored in the Huggingface LeRobot dataset format
• Model training with dual verification: NVIDIA Isaac GR00T N1.6 is fine-tuned on NVIDIA DGX, and the trained model undergoes a Model Verification (V&V) process on NVIDIA OVX — only models that pass verification are deployed to the physical robot
• NVIDIA Jetson Thor edge deployment: Verified models are deployed on the NexCOBOT MARS400 Pro platform, powered by NVIDIA Jetson Thor, which converts RGBD images into joint control commands in real time via a ROS2 node to drive the robot arm
• The technology was demonstrated live at the 2026 TAIROS show, successfully showcasing the complete Real2Sim2Real loop and the robot's ability to identify, grasp, and sort specific objects.
  
  
  

From Human Demonstration to Large-Scale Simulation: Data Collection and Generation

One of the biggest challenges in robot motion learning is the high cost and limited quantity of real-world demonstration data. ITRI uses low-latency teleoperation to collect precise human demonstration motion trajectories as ground-truth data for Imitation Learning. A digital twin is then built using NVIDIA Omniverse and Isaac Sim, combined with NVIDIA Cosmos and Mimic/Gen technology, to scale a small amount of real demonstration data into a large-scale, highly randomized virtual training dataset — significantly reducing the time and cost of data collection while improving the model’s ability to generalize across varied real-world scenarios.

Training and Verification: The Dual Safeguards of DGX, GR00T N1.6, and OVX

The collected and generated datasets are stored in the Huggingface LeRobot dataset format, and NVIDIA Isaac GR00T N1.6 is fine-tuned on NVIDIA DGX to train a VLA model with vision, language, and motion understanding capabilities. Rather than testing the trained model directly on hardware, it first undergoes a Model Verification (V&V) process on NVIDIA OVX: the model performs inference validation against designated test tasks within Isaac Sim, and only models that pass verification are deployed to the physical robot — ensuring the stability and safety of the on-robot model.

NexCOBOT MARS400 Pro with NVIDIA Jetson Thor: Bringing the AI Controller to Real-World Execution

The verified model is ultimately deployed on the NexCOBOT MARS400 Pro platform — powered by the latest NVIDIA Jetson Thor compute core, serving as the edge computing hub for on-site robot execution. The platform receives RGBD images and task lists in real time via a ROS2 node, runs inference through the post-trained model to output joint control commands (Control Law) that drive the robot arm, while simultaneously receiving real-time Arm Status feedback to form a closed-loop control system. This “real-time AI controller” architecture enables the robot to demonstrate stable, autonomous judgment and manipulation in complex, unstructured real-world settings — such as the pick-and-sort tasks demonstrated live at the show — representing the critical last mile of the Real2Sim2Real technology chain.

About the Industrial Technology Research Institute (ITRI)

ITRI is an international applied research institution with more than 6,500 research professionals, dedicated to driving industrial development, creating economic value, and enhancing social well-being through technology R&D. Since its founding in 1973, ITRI has pioneered integrated circuit R&D and nurtured emerging technology industries, accumulating more than 30,000 patents and incubating publicly listed companies including TSMC, UMC, Taiwan Mask Corporation, Epistar, Mirle Automation, and TSMBC, driving successive waves of industrial development.

As digital technology reshapes industrial structures and lifestyles, demographic shifts influence productivity and elder-care needs, climate change brings both opportunities and challenges on the path to net-zero emissions by 2050, and industrial and social resilience become key trends for national and economic development, ITRI integrates cross-domain solutions to accelerate industrial momentum, charting mid- to long-term technology strategies and roadmaps. ITRI focuses its R&D on four core application domains — Smart Living, Quality Health, Sustainable Environment, and Resilient Society — while developing Smart-Enabling Technologies to support these domains. Through technology-driven innovation, ITRI transforms everyday life, developing market-driven solutions and creating new markets to advance human well-being and lead industry and society toward a better future. Learn more at itri.org.tw.

About NexCOBOT

NexCOBOT is a technology provider in robot control and motion control — not a robot manufacturer, but the computing and control platform that enables robots to sense, think, and move safely and intelligently. With more than 13 years of robotics industry experience, NexCOBOT integrates AI computing, real-time motion control, and industrial functional safety into a single architecture, working closely with NVIDIA to provide humanoid robots, quadruped robots, collaborative robots, and industrial AI systems with a complete platform spanning computing, control, and safety. Learn more at nexcobot.com.

This press release contains forward-looking statements. Actual results may differ materially from those statements due to market conditions, technological developments, or other factors.