NVIDIA has released its latest Vision-Language-Action (VLA) foundation model, GR00T 1.7, on the LeRobot platform. This model is designed to enhance open robot learning and humanoid development, offering improved performance and expanded manipulation capabilities over its predecessor, GR00T N1.5. Developers are encouraged to transition to GR00T 1.7 as N1.5 is no longer supported. The model is also available on the NVIDIA Isaac GR00T open development platform, which includes pre-integration with the GR00T reference humanoid robot for academic research. The release aims to streamline the workflow from data collection to imitation learning and policy fine-tuning by providing a seamless path from development to deployment.
The GR00T 1.7 model allows developers to quickly adapt it to their own tasks and robots using LeRobot's open workflows. Models trained anywhere can be deployed through LeRobot while maintaining benchmarked performance, ensuring consistency across different environments. The release includes a guide for installation, training, and deployment with GR00T 1.7, which walks users through the process of executing a real-world task using the SO-101 robot. The guide outlines the necessary commands for setting up the environment and collecting data via either a VR headset or the SO-101 leader arm, ensuring compatibility with downstream training pipelines.
The source text describes the availability of GR00T 1.7 on LeRobot, emphasizing its role in accelerating end-to-end humanoid development and providing a seamless path from development to deployment. It outlines the steps for data collection using Isaac Teleop, including the setup of environment variables and the commands for running the teleoperation framework. The text also mentions the use of a RTX 6000 Pro for testing and the process for training and deploying the model using LeRobot's tools. The release includes instructions for logging into Hugging Face and using the wandb login for tracking training progress.
Source: huggingface
