AutoLab-SJTU/robofollow-data

RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents This dataset is part of the RoboFollow benchmark, designed to evaluate instruction-following capabilities of embodied agents in manipulation tasks. Built on RoboTwin, it provides shared-scene task configurations where multiple tasks coexist, requiring policies to rely on language to disambiguate objectives. The dataset includes expert demonstrations and evaluation tasks across four scene families, covering… See the full description on the dataset page: huggingface.co/datasets/AutoLab-SJTU/robofollow-data.