This repository contains the officially formatted LeRobot (v3.0+) datasets for the LIBERO benchmarks, augmented via physical state perturbation.
Training robust offline reinforcement learning (RL) or Vision-Language-Action (VLA) models often suffers from the narrow distribution of expert demonstrations. This dataset aims to artificially inflate the robustness of offline datasets by providing… See the full description on the dataset page:
https://huggingface.co/datasets/jdhr/libero_trajectory_augmented.
