# Spotlight: Fourier Trains Humanoid Robots for Real-World Roles Using NVIDIA Isaac Gym *Published:* 2024-11-06 *Author:* Ben Oliveri ## AI-generated summary - Fourier, a Shanghai-based robotics company, has launched GR-2, an upgraded humanoid robot with advanced dexterity and humanlike range of motion, building on its previous GR-1 model. - The development of GR-2 was supported by NVIDIA Isaac Gym, which allowed the Fourier team to simulate real-world conditions, minimizing testing time and cost, and is now being ported to NVIDIA Isaac Lab. - By utilizing NVIDIA technologies such as NVIDIA TensorRT and CUDA libraries, Fourier significantly reduced model training times and improved simulation accuracy, enabling complex AI functions like language models and predictive analytics. *This post was written in partnership with the* [*Fourier*](https://www.fftai.com/) *research team.* Training humanoid robots to operate in fields that demand high levels of interaction and adaptability—such as scientific research, healthcare and manufacturing—can be a challenging and resource-intensive feat. [Fourier](http://www.fftai.com/), a Shanghai-based robotics company, is doing the heavy lifting by developing advanced humanoid robots that can be integrated into real-world applications where precision and agility are critical. The company announced the expansion of its GRx humanoid robot series with the launch of GR-2 in late September. Building on the previous-generation GR-1, the world’s first mass-produced humanoid robot, GR-2 features an upgraded hardware design, greater adaptability, advanced dexterity and a humanlike range of motion. Developing humanoid robot GR-2 with NVIDIA Isaac Gym[](#developing_humanoid_robot_gr-2_with_nvidia_isaac_gym) ------------------------------------------------------------------------------------------------------------- To develop and test GR-2, the Fourier team turned to NVIDIA Isaac Gym (now deprecated) for reinforcement learning. They are currently porting their workflows to the recently launched [NVIDIA Isaac Lab](https://developer.nvidia.com/isaac/lab), an open-source modular framework for robot learning designed to simplify how robots adapt to new skills. Sim-to-real learning has become essential for robotics, especially for complex movements like sitting down, getting up, or even dancing. With Isaac Gym, Fourier was able to simulate real-world conditions, minimizing the time and cost of testing and maintenance. The team simulated complex multi-robot scenarios and real-world environments, leading to more robust AI decision-making and enhanced real-world performance—even in unpredictable settings. Fourier also used Isaac Gym to pretrain grasping algorithms, simulating success rates before deployment. This approach significantly reduces the real-world trial and error, saving time and resources. **Optimizing AI for real-world robotics**[](#optimizing_ai_for_real-world_robotics) ----------------------------------------------------------------------------------- While training GR-2 for the floor-to-stand maneuver, Fourier simulated the physical demands required for completing tasks at different levels of elevation. By replicating the GR-2 model, they tested how it performs under various settings and completed 3,000 iterations in around 15 hours, a notable reduction compared to traditional training methods. When transferred directly to GR-2’s physical controls, the model’s action tensors achieved an 89% success rate. ![Robots perform the floor-to-stand maneuver in an animated image at a checkpoint after 100 test iterations.](https://developer-blogs.nvidia.com/wp-content/uploads/2024/11/Checkpoint100-1.gif) *Checkpoint after 100 test iterations*![Robots perform the floor-to-stand maneuver in an animated image at a checkpoint after 500 test iterations.](https://developer-blogs.nvidia.com/wp-content/uploads/2024/11/Checkpoint500-2.gif) *Checkpoint after 500 test iterations*![Robots perform the floor-to-stand maneuver in an animated image at a checkpoint after 1,600 test iterations.](https://developer-blogs.nvidia.com/wp-content/uploads/2024/11/Checkpoint1600-1.gif) *Checkpoint after 1,600 test iterations*![Robots perform the floor-to-stand maneuver in an animated image at a checkpoint after 3,000 test iterations.](https://developer-blogs.nvidia.com/wp-content/uploads/2024/11/Checkpoint3000-1.gif) *Checkpoint after 3,000 test iterations**Figure 1. The Fourier team observed a significant increase in performance success rates of the floor-to-stand maneuver after conducting 1,600 test iterations*To enhance the development process, the team tapped the[ NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) software development kit for real-time inference optimization, CUDA libraries for parallel processing, and the [NVIDIA cuDNN](https://developer.nvidia.com/cudnn) library for accelerating deep learning frameworks like PyTorch. Moving to [NVIDIA Isaac Lab](https://developer.nvidia.com/isaac/lab) will enable Fourier to train more complex algorithms and simulations in multiphysics virtual environments powered by NVIDIA RTX tiled rendering. Exploring next-generation robotic capabilities[](#exploring_next-generation_robotic_capabilities) ------------------------------------------------------------------------------------------------- By adopting NVIDIA technologies, Fourier significantly reduced model training times and improved the accuracy of simulations, which resulted in enhanced collaboration across its engineering and R&D teams. NVIDIA tools also opened the door to complex AI functions like language models and predictive analytics, previously too resource-heavy to implement. “The advancements we’ve achieved are pushing the boundaries of what’s possible in humanoid robotics,” said Fourier CEO Alex Gu. “By improving the robot’s real-time motion control and AI-driven decision-making, we are setting new standards for human-robot interaction across industries such as the service sector, academic research, and medical rehabilitation.” [Learn more about Fourier GR-2 humanoid robots](http://www.fftai.com/). Get started[](#get_started) --------------------------- Need to migrate from NVIDIA Isaac Gym to [NVIDIA Isaac Lab](https://developer.nvidia.com/isaac/lab)? Check out the [Isaac Lab Migration Guide](https://isaac-sim.github.io/IsaacLab/main/source/migration/migrating_from_isaacgymenvs.html). If you’re a first-time user of Isaac Lab, see the [Getting Started Developer’s Guide](https://isaac-sim.github.io/IsaacLab/main/source/tutorials/index.html#). Discover the latest in robot learning and simulation in the [November 13 livestream, OpenUSD Insider Livestream – Robot Sim and Learning](https://www.addevent.com/event/GA23422424). And don’t miss the [NVIDIA Isaac Lab Office Hours](https://www.addevent.com/event/Uz23738360) for hands-on support and insights.