How to train generalist robots with NVIDIA’s research workflows and foundation models

NVIDIA researchers are developing methods to train robots without extensive data collection for each task. Using generative AI and foundation models like Cosmos, they've created systems including DreamGen (which generates synthetic training data), GR00T N1 (an open foundation model for humanoid robots), and techniques for learning from unlabeled videos. These approaches can reduce development time from months to hours while improving robot performance across diverse tasks and environments.