Quartz 5

AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems

Properties1
tagsclippings, robotics, dataset, teleoperation, manipulation, humanoid, large-scale

2 min read

AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems

arXiv: 2503.06669 | Venue: IROS 2025 Best Paper Finalist + IEEE TRO 2026

Authors: OpenDriveLab / AgiBot team

Abstract

AgiBot World is a large-scale platform comprising over 1 million trajectories across 217 tasks in five deployment scenarios. The platform introduces Genie Operator-1 (GO-1), a generalist policy using latent action representations, trained on this dataset. Policies trained on AgiBot World show 30% performance improvement over those trained on Open X-Embodiment and over 60% success rate on complex tasks.

Dataset Scale

  • Trajectories: 1,001,552+ (AgiBot World Beta: 1,003,672; Alpha subset: 92,214)
  • Total duration: 2,976 hours
  • Tasks: 217 specific tasks
  • Skills: 87
  • Scenes: 106
  • Deployment scenarios: 5

Hardware & Collection

  • Robot type: Humanoid robots with grippers (extensible to dexterous hands + visuo-tactile sensors)
  • Teleoperation: Skilled teleoperators with human-in-the-loop verification
  • Quality control: Standardized collection pipeline with multi-stage human verification

Key Contributions

  1. Scale: Order-of-magnitude increase over prior datasets (1M+ vs. DROID’s 76k, RH20T’s 110k)
  2. Quality control: Rigorous multi-stage human verification ensures high-quality diverse distribution
  3. GO-1 policy: Latent action representation-based generalist policy achieving SOTA on complex tasks
  4. Performance: 30% better than Open-X pretrained policies; 32% better than RDT approach
  5. Open release: Data available on HuggingFace, GitHub, OpenDataLab

Tiered Release

  • Alpha subset: 92,214 trajectories (~8.5T tokens) — curated high-quality subset
  • Beta full: 1,003,672 trajectories (~43.8T tokens) — full dataset

Context

AgiBot World represents the Chinese robotics ecosystem’s entry into the large-scale pretraining data race, analogous to how AgiBot (formerly part of Zhiyuan Robotics) is positioning itself against US-based datasets. The IROS 2025 Best Paper Finalist recognition validates the dataset’s research impact.

Graph View

Backlinks