NVIDIA and Hugging Face Open Up Tools for Humanoid Robot Development

NVIDIA and Hugging Face announced an expansion of the open LeRobot library with tools for humanoid and other advanced robots. They added the Isaac GR00T 1.7 model and the Isaac Teleop framework to the shared environment. The Cosmos 3 physical world model is expected to follow later. The aim is to connect data collection, model training, simulation, validation and robot deployment in one standardised workflow. According to the companies, robotics development is currently constrained by the high cost of data, fragmented software and difficult access to computing power. Teams often create proprietary formats and tools that are hard to share. LeRobot is used to manage robot data, models, behavioural policies and workflows. Hugging Face reports a community of 16 million artificial intelligence developers. NVIDIA says it has three million developers focused on robotics. Their combination could broaden the group of companies able to develop robots without building their own foundation model. Isaac Teleop allows developers to record human demonstrations of robot control through external devices. A developer can demonstrate how to grasp an object, manipulate a tool or move through a workspace. The recording is stored in a shared format and can be expanded or shared with other teams. GR00T 1.7 is an open model that combines images, language and movement commands. NVIDIA describes it as the first open and commercially usable foundation model for robots of this type. Through LeRobot, companies can adapt it to specific machines and work tasks. “Open access turns advanced research into something people can study and modify,” said Hugging Face co-founder Thomas Wolf. In his view, shared models, data and workflows will allow developers to train and evaluate robots more openly. The open environment includes a dataset downloaded more than fifteen million times. It contains over 350,000 real and simulated trajectories and 57 million attempts to grasp an object. The dataset is intended to let developers start from existing material rather than record everything from scratch. NVIDIA is also connecting LeRobot with the Isaac Sim and Isaac Lab simulators. Before deployment on a real machine, teams can create environments, test behaviour and identify faulty reactions. Isaac Lab-Arena supports more complex simulation scenarios and comparison of different general-purpose models. For industrial companies, standardisation may reduce testing, integration and supplier-switching costs. Openness, however, does not remove responsibility for safety. A model controlling a physical device can cause damage through an incorrect movement, failure to recognise a person or an inappropriate response to an unexpected object. A company therefore needs to separate the research environment from the production network, limit robot speed and force, and install an independent emergency stop. Every model change must be validated again. In Europe, product liability will be supplemented by machinery rules, employee protection requirements and, depending on use, artificial intelligence regulation.

NVIDIA and Hugging Face Open Up Tools for Humanoid Robot Development

The full article is available to users registered at Hard Skills