Yassine/ogbench-block-double-hermite-100k

ogbench-block-double-hermite-100k Unofficial reproduction of the scripted policies described in https://seohong.me/blog/behavioral-cloning-mystery/ using random piecewise Hermite splines as the backbone, and randomized control points, grasping angles/directions, contact points, motion speed, gripper yaw/roll/pitch, mistakes and retries, etc. This might not be the exact setup used by the study, but I tried to infer the parameters from "How exactly did you script the policies?"… See the full description on the dataset page: https://huggingface.co/datasets/Yassine/ogbench-block-double-hermite-100k.