GR00T N1

We post-train GR00T N1 and N1.5 on 1K teleoperation episodes collected on the Unitree G1 robot. As in the GR-1 language following experiment, we initialize the scene with one target object and one distractor object, with the target object having equal probability of being closer to either the left or the right hand. We observe that the post-trained GR00T N1.5 achieves much higher success rate than N1 for previously seen objects (toy fruits seen in the GR-1 pretraining corpus), and also shows generalization to various previously unseen objects.