RealWorld Debuts Robot Hand AI Model

RealWorld's new RLDX-1 foundation model gives humanoid robot hands advanced dexterity, built entirely on NVIDIA's cloud-to-edge AI and robotics platforms.

RealWorld has officially unveiled RLDX-1, a new foundation model designed to provide sophisticated manipulation capabilities for five-fingered humanoid robot hands. Introduced at NVIDIA GTC Taipei, the "Dexterity-First" model aims to solve one of the most complex challenges in robotics: fine motor control. The development of RLDX-1 relies heavily on NVIDIA's core robotics development platforms. RealWorld utilized a suite of tools including Isaac GR00T, a general-purpose foundation model for humanoid robots, alongside Isaac Lab for reinforcement learning, Isaac Sim for realistic simulation, and cuRobo for motion planning.

The significance of RLDX-1 lies in its complete integration with NVIDIA's cloud-to-edge computing stack, offering a clear blueprint for developers and CTOs in the AI and robotics sectors. For the intensive training and development phases in the cloud, RealWorld leverages the power of NVIDIA H100 and A100 Tensor Core GPUs. When the model is deployed on the robot itself, it runs on NVIDIA's edge computing platforms, specifically the Jetson AGX Thor and Jetson Orin systems. To ensure optimal performance on these edge devices, the model is optimized using NVIDIA TensorRT. This end-to-end implementation showcases a scalable pathway for building advanced robotics applications using a unified ecosystem.

The project serves as a key proof-of-concept for building complex, end-to-end robotics applications entirely within the NVIDIA ecosystem, from cloud training on H100 GPUs to edge deployment on Jetson hardware.

Companies in the robotics and automation sectors can see a viable, unified platform for developing and deploying advanced manipulation capabilities, potentially accelerating product development and reducing reliance on disparate technology stacks.