AgiBot World
Open manipulation data and policy weights from AgiBot (智元机器人): a million-plus real robot trajectories, the GO-1 VLA and the Genie Envisioner world model, all non-commercial.
Overview
AgiBot, 智元机器人 in Chinese, runs its open programme through OpenDriveLab, and the release has three separate pieces that people often confuse. There is AgiBot World, the dataset. There is GO-1, the policy trained on it. And there is Genie Envisioner, a world model platform with its own simulator and benchmark. The Beta dataset holds 1,003,672 real robot trajectories collected across more than a hundred environments rebuilt one to one from shops, homes and offices, with visuo-tactile sensing and six-DoF dexterous hands in the rig rather than a single gripper.
The April 2026 drop, AgiBot World 2026, moved to the G2 platform and landed at 9.36 TB over 217 tasks, staged as the first of five planned phases each aimed at a different research direction. It feeds the AgiBot World Challenge at ICRA 2026, which added Reasoning to Action and World Model tracks. GO-1 itself is small by language model standards, 3B parameters on an InternVL 2.5-2B backbone, and the interesting part is the split between a latent planner trained on cross-embodiment and human video and an action expert trained on the robot demonstrations. GO-1 Air removes the planner for teams who want the throughput and can live with less generalisation.
Practicalities matter here more than benchmark tables. Weights and data are mirrored across Hugging Face and ModelScope, which is the right call for a Chinese lab and rare enough to be worth praising, though a few Genie Envisioner checkpoints exist only on ModelScope. Inference sits near 7 GB of VRAM, a full fine-tune wants an 80 GB card, and the action-expert path fits a 4090. The blocker is legal, not technical. CC BY-NC-SA 4.0 means a research lab or a thesis, not a shipped product, and any team planning commercial work should be talking to AgiBot about terms before a single epoch runs.
Key Features
- ✓ AgiBot World 2026 dataset, released April 2026 at 9.36 TB across 217 tasks and five deployment scenarios, recorded on the AGIBOT G2 dual-arm platform. The Hugging Face repo has since grown past 13 TB.
- ✓ AgiBot World Beta: 1,003,672 real robot trajectories from more than 100 scenes replicated one to one, captured with visuo-tactile sensors, six-DoF dexterous hands and mobile dual-arm rigs.
- ✓ GO-1 (Genie Operator-1), a 3B vision-language-action model on an InternVL 2.5-2B backbone using the ViLLA scheme, a latent planner plus an action expert in a mixture of experts. GO-1 Air, released 19 September 2025, drops the latent planner for lower latency.
- ✓ Failed demonstrations are kept and annotated with error_cause and restorable fields rather than filtered out, which makes the corpus usable for recovery policies and reward models.
- ✓ Genie Envisioner: GE-Base video world model, GE-Act action head, GE-Sim (v2.0 announced May 2026) and the EWMBench embodied world model benchmark.
- ✓ Published VRAM figures: about 7 GB for inference on an RTX 4090, about 70 GB at batch 16 for a full fine-tune on an A100 80GB or H100, about 24 GB if you fine-tune the action expert only.
Where it holds
- • Real robot data at a scale nobody else gives away, with hardware and sensor metadata detailed enough for cross-embodiment training.
- • Mirrored on both Hugging Face and ModelScope, so the download works from inside the mainland without a tunnel.
- • Action-expert-only fine-tuning fits in roughly 24 GB, which puts a working policy inside reach of one consumer GPU.
- • Failure trajectories are labelled instead of discarded, a design choice almost no other open manipulation dataset makes.
Where it breaks
- • CC BY-NC-SA 4.0. Non-commercial and share-alike, applying to the weights as well as the data, so anything you fine-tune from GO-1 inherits the restriction.
- • Several checkpoints, including the Calvin-trained GE-Act weights and the GE-Sim Cosmos build, are on ModelScope only, which is slow and account-gated from outside China.
- • Everything is recorded on AgiBot's own G1 and G2 hardware. Action spaces and camera extrinsics need remapping before the data means anything on your robot.
- • READMEs are in English but the issue threads and much of the surrounding discussion are Chinese-first, so debugging without Mandarin is slower than the docs suggest.
My Take
Start with the licence, because it decides everything else. CC BY-NC-SA 4.0 covers the dataset and the GO-1 weights alike, so this is research fuel rather than a product ingredient, and share-alike follows anything you fine-tune from it. What you get for accepting that is scale nobody else offers free: 9.36 TB in the April 2026 release across 217 tasks, with failed grasps kept and tagged by error cause instead of quietly deleted. Fine-tuning the action expert alone runs in about 24 GB, so a single RTX 4090 is enough to get to a policy worth testing.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- CC BY-NC-SA 4.0 (dataset, GO-1 weights and most Genie Envisioner code); Apache 2.0 for the borrowed LTX, Cosmos and openpi client components inside Genie Envisioner
- Starting at:
- Free to download. The dataset, the GO-1 and GO-1 Air checkpoints and the Genie Envisioner weights cost nothing. The constraint is the licence, not the price: CC BY-NC-SA 4.0 rules out commercial deployment, and a commercial arrangement has to be negotiated with AgiBot directly.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where AgiBot World holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to AgiBot World
RoboBrain 2.5
open-sourceBAAI's embodied brain model (悟界·RoboBrain), Apache 2.0 at 4B and 8B, shipped in parallel NVIDIA and Moore Threads builds. It plans and points in 3D for a controller underneath.
Unitree UnifoLM
open-sourceUnitree's open robot learning stack: world models, a vision-language-action policy, Isaac and MuJoCo training environments and the SDK that drives Go2, G1 and R1 hardware.
Unitree Robotics Developer Stack
open-sourceThe open repos behind Unitree's robots (宇树科技): BSD-3 C++ and Python SDKs, MuJoCo and IsaacLab RL training, ROS 2 packages and XR teleoperation for data capture.