Nvidia used Computex 2025 to push its humanoid robot strategy further into practical industrial work. The company unveiled GR00T N1.5, an updated foundation model designed to help humanoid robots understand new settings, recognize material handling tasks, and perform them with more general capability.
The release builds on GR00T N1, which was announced in March. The core idea is clear: give robots a broader base of reasoning and action skills, then use synthetic data and simulation to reduce the time and effort needed to train them.
GR00T N1.5 targets flexible industrial work
GR00T N1.5 is aimed at one of robotics' hardest practical problems: getting machines to work reliably outside tightly controlled routines. Nvidia says the new model is designed to better handle unfamiliar environments and can recognize and perform a wide range of material handling tasks.
That matters because many industrial settings depend on repeated movement, object handling, and adaptation to changing layouts or workflows. A robot that can only perform a narrow programmed motion is useful, but limited. A humanoid robot that can interpret a task and act in a new environment could be applied more broadly.
The model uses a dual-system architecture. System 2 is responsible for cognitive tasks such as planning, while System 1 controls real-time motor execution. In plain terms, one part of the system helps decide what should happen, and the other manages the physical movement needed to do it.
Nvidia frames this as a step toward giving humanoid robots generalized reasoning and action skills. The comparison in the source is to the way large language models have changed language tasks: the ambition is not just to program one behavior, but to build a foundation that can support many behaviors.
Synthetic data changes the training timeline
One of the most important claims around GR00T N1.5 is the speed of training. According to Nvidia, the model can be trained in just 36 hours using synthetic data. The same process would normally take about three months with traditional methods.
That difference explains why synthetic data is central to the announcement. Real-world robot training depends on collecting physical examples, and the source notes that real-world robots can collect only a limited amount of data each day. If training data can be generated instead, developers can scale the learning process without waiting for robots to repeatedly perform tasks in the physical world.
Early adopters named in the source include AeiRobot, Foxlink, Lightwheel, and NEURA Robotics. Their use cases point toward manufacturing and industrial automation rather than consumer robotics.
- AeiRobot is using the model to control industrial pick-and-place workflows through natural language.
- Foxlink aims to make its industrial robots more flexible.
- Lightwheel is validating synthetic training data for humanoid robots in manufacturing.
- NEURA Robotics is also listed among the early adopters.
These examples show the near-term focus of GR00T N1.5. The model is not being presented as a general household assistant. It is being positioned around work where physical handling, repeatability, and adaptability are valuable.
GR00T-Dreams turns images into robot motion data
Alongside GR00T N1.5, Nvidia introduced GR00T-Dreams, a blueprint for producing synthetic motion data. The system uses image-driven AI video models to generate training examples for robots.
GR00T-Dreams was developed under the direction of Jim Fan, who leads Nvidia's research group for embodied generative AI. Its workflow begins with developers training a world model using Cosmos Predict. Then GR00T-Dreams takes a single image and generates a video of a robot performing a new task in a new environment.
From that video, the system extracts action tokens. The source describes these as compressed data fragments that teach the robot new behaviors. Instead of relying only on physical demonstrations, developers can create synthetic examples and feed them into the training pipeline.
The videos are produced by the Cosmos AI video generator, which has been fine-tuned with robot footage from Nvidia's labs. That detail is important because the videos are not just general AI clips. They are part of a training system shaped around robot motion and robot behavior.
The practical advantage is scale. If a robot can only gather so much data in a day, training progress is limited by time and physical operation. GR00T-Dreams is designed to let developers generate as much training data as needed.
Simulation tools aim to reduce the sim-to-real gap
Nvidia also announced updates to simulation and data frameworks that support training and testing. These include Isaac Sim 5.0 and Isaac Lab 2.2, both available on GitHub.
Isaac Lab now includes new test environments for GR00T N models. Nvidia is also releasing an open-source dataset with 24,000 high-quality motion sequences for humanoid robots. Together, these tools are meant to help developers test, refine, and expand robot behavior before or alongside real-world deployment.
The company also made a new world model called Cosmos Reason available. It uses a "chain of thought" approach to curate high-quality synthetic training data. Cosmos Predict 2, which powers GR00T-Dreams, is set to launch soon on Hugging Face.
Another part of the system is GR00T-Mimic, a blueprint introduced in March. It generates large volumes of synthetic motion data for manipulation tasks from just a handful of human demonstrations. Foxconn and Foxlink are already using it to speed up their training pipelines.
The bigger shift is from programming to training
The announcements point to a broader change in humanoid robotics. Nvidia is building a stack in which models, synthetic videos, action tokens, simulation environments, and open datasets work together to accelerate training.
For developers, the appeal is not only that robots may learn more behaviors. It is that the process of teaching those behaviors may become faster and less dependent on physical data collection. GR00T N1.5, GR00T-Dreams, Cosmos Predict, Cosmos Reason, Isaac Sim 5.0, Isaac Lab 2.2, and GR00T-Mimic all fit into that same direction.
The focus remains practical. The named examples are industrial pick-and-place workflows, flexible industrial robots, humanoid robots in manufacturing, and manipulation tasks. Nvidia's humanoid robot work is therefore less about science fiction and more about shortening the path from model training to factory use.