MolmoAct Dataset
Allen Institute for AI (Ai2) · 2026 · datasets.bot · datasets.bot page
One-liner. Ai2's own MolmoAct real-robot Franka dataset: household + tabletop EE manipulation, 3 camera views.
Setup
- Datasets / benchmarks: Ai2's self-collected MolmoAct real-robot manipulation dataset, used to train MolmoAct action reasoning models, covering household and tabletop domains. The Household split has 7,529 episodes across 115 tasks with varying viewpoints; the Tabletop split adds 2,959 episodes across 19 tasks with a fixed exocentric camera (~10,488 episodes total). It uses a Franka arm with three RGB views (primary, secondary, wrist) and a 7-dim end-effector action space, in LeRobot format with per-episode language annotations. License: Apache-2.0. Download: https://huggingface.co/datasets/allenai/MolmoAct2-MolmoAct-Dataset-Household.
- Hardware / simulator: Embodiment: franka_panda. Environment: home, tabletop. Realness: physical.
Schema
LeRobot: 3 RGB videos (primary/secondary/wrist, 480x640) + 7-dim EE state and 7-dim absolute/delta EE actions; per-episode language instructions. Household + Tabletop splits.
Links