A Million Robot Trajectories and a Model to Boot

What it does

AgiBot World Colosseo is a full-stack platform for bimanual robot manipulation research. It packages a dataset of over one million trajectories from 100 robots with a pretrained vision-language-action foundation model called GO-1 and benchmarks built on the LeRobot framework. The aim is to provide a common starting point for generalist robot policies, pitched as an “ImageNet Moment” for embodied AI.

The interesting bit

Rather than dumping raw logs and leaving, the team released the entire ecosystem: two model variants (GO-1 with a Latent Planner, and the lighter GO-1 Air), a task catalog, and conversion tools to standardize external datasets into LeRobot format. It is essentially a bet that robotics needs a shared data and model backbone, not just scattered downloads.

Key highlights

• Scale: The Beta dataset holds 1,003,672 trajectories (~43.8 TB) gathered from 100 robots across more than 100 replicated real-world scenarios.
• Pretrained policies: GO-1 and GO-1 Air checkpoints are hosted on HuggingFace, ready for fine-tuning on custom manipulation tasks.
• Hardware diversity: Data includes mobile dual-arm robots, 6-DoF dexterous hands, and visual tactile sensors.
• Academic recognition: IROS 2025 Best Paper Award Finalist and accepted to IEEE TRO 2026.
• Non-commercial license: Everything is under CC BY-NC-SA 4.0, which keeps it in the research realm.

Caveats

• The README is mostly quickstart scripts and installation; the actual architecture and design rationale live in the separate technical report.
• Full fine-tuning requires roughly 70 GB of GPU memory, pushing most users toward the lighter GO-1 Air or partial fine-tuning modes.
• The CC BY-NC-SA 4.0 license means commercial deployment is off the table unless you negotiate separately.

Verdict

This is a goldmine for academic robotics labs working on generalist manipulation and long-horizon planning. Industry teams or hobbyists without A100-class hardware should admire from a distance.