AgiBot World: Large-Scale Multi‑Robot Embodied AI Dataset Release

The AgiBot World dataset, released by Zhiyuan Robotics in collaboration with Shanghai AI Lab and other partners, is the first globally‑scale robot dataset captured in fully realistic environments with end‑to‑end quality control.

Compared with existing datasets such as DeepMind’s Open X‑Embodiment and the DROID dataset, AgiBot World offers ten times the trajectory length and one hundred times the scene coverage, moving from laboratory‑grade to industrial‑grade data quality.

The dataset contains over 80 diverse daily‑life skills, ranging from basic grasp‑and‑place actions to fine‑grained long‑range two‑arm collaborations such as memory‑chip insertion, dishwasher organization, ironing, and cooperative object transport.

Data were collected in a 4,000 m² facility covering more than 3,000 real objects and over 100 real‑world scenes. The distribution of scenarios includes Home (40%), Dining (20%), Industrial (20%), Retail (10%) and Office (10%).

All recordings were captured by a custom 32‑DOF robot equipped with 360° vision, a 6‑DOF dexterous hand, six‑dimensional force sensing, and optional high‑precision tactile sensors, ensuring precise and safe manipulation.

Quality assurance involves multi‑stage pipelines: task design reviewed by academia and industry, professional operator training, automated filtering, frame‑by‑frame manual verification, and algorithmic validation. Low‑quality samples are re‑collected.

Future releases will provide the full million‑robot‑run dataset, tens of millions of simulated trajectories, a foundational embodied AI model, a complete toolchain for data collection‑to‑evaluation, and a series of AgiBot World challenges.

Project links: GitHub , HuggingFace , Official site .