metadata
task_categories:
- robotics
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 spatial relations, object attributes, trajectory constraints, and conditional instructions. It supports the benchmark's hierarchical evaluation protocol (L0–L3) to separate intent comprehension from physical execution.
For more details, refer to:
- Paper: arXiv:2609.25636
- Project page: RoboFollow Project
- Code: GitHub Repository