AY-ROBOTSAI Teleoperation
Sign Up
AGI

GO-1 (Genie Operator-1)

AgiBot, China · March 2025

AgiBot (Zhiyuan Robotics) with OpenDriveLab

Foundation VLAOpen weights, CC BY-NC-SAAlso written Genie Operator-1, AgiBot GO-1
Parameters
3 B
The Hugging Face model card states 3B parameters, BF16.
GPU memory
5.6 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
CC BY-NC-SA
CC BY-NC-SA 4.0 (non-commercial, share-alike)

What it is

GO-1 is AgiBot's generalist manipulation policy and the model component of the AgiBot World Colosseo release. Its distinguishing idea is ViLLA: instead of mapping vision and language straight to actions, an InternVL2.5-2B backbone feeds a 24-layer Latent Planner that predicts four discretized latent action tokens, which a diffusion Action Expert then decodes into a 30-step action chunk at 30 Hz. The point of the latent layer is data leverage, since latent actions can be learned from action-free human video such as Ego4D and then grounded on the 1M-plus trajectory AgiBot World dataset. The paper reports 0.77 versus 0.47 task completion in seen scenarios against Open X-Embodiment pre-training and a power-law data-scaling fit with r = 0.97. The blocking issue for commercial users is the CC BY-NC-SA 4.0 licence on the weights, and no VRAM or latency figures are published at all.

Architecture

Backbone
InternVL2.5-2B
Action head
Mixture of experts in the ViLLA (Vision-Language-Latent-Action) framework: a Latent Planner of 24 transformer layers with full bidirectional attention that predicts k=4 discretized latent action tokens from a VQ-VAE codebook, followed by an Action Expert that decodes continuous low-level actions with a diffusion objective
Action chunk
30 steps per forward pass
Control rate
30 Hz
Parameters
The Hugging Face model card states 3B parameters, BF16. The vision-language backbone is InternVL2.5-2B, so roughly 1B sits in the latent planner and the diffusion action expert on top of it. The paper gives the latent planner as 24 transformer layers but does not break out parameter counts per component.
Pretraining data
Three-stage curriculum. Stage 1 trains the latent action model on internet-scale heterogeneous video including Ego4D human video. Stage 2 trains the Latent Planner on AgiBot World. Stage 3 jointly trains the Action Expert on AgiBot World, a platform of over 1 million trajectories across 217 tasks in five deployment scenarios. The Hugging Face card references AgiBotWorld-Beta (236k samples) and AgiBotWorld-Alpha (16k samples).
Embodiments
AgiBot G1 (bimanual mobile manipulator), grippers, dexterous hands and visuo-tactile sensors per the platform description

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 3 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3211 GBTraining master weights, rarely used for inference
bf16 or fp165.6 GBThe usual way these checkpoints are served
int82.8 GBQuantised, expect some loss of precision on fine motions
int41.4 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Control rate30 Hz

No VRAM or GPU requirement is published on the Hugging Face card or in the paper. At 3B parameters in BF16 the weights alone are roughly 6GB, but no source states a usable inference or fine-tuning figure, so none is recorded here.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot (complex long-horizon and dexterous tasks) aggregate over the paper's complex task set
    60%
    success rateAuthors' own evaluation, stated in the abstract as 'over 60 percent' and as outperforming RDT by 32 percent. No per-task breakdown is given in the abstract.source
  • AgiBot World (real robot, AgiBot hardware) in-domain / seen scenarios, pre-trained on AgiBot World vs on Open X-Embodiment
    0.77 score (0 to 1)
    average task completion scoreAuthors' own evaluation. The OXE-pretrained comparison policy scores 0.47 in the same setting.source
  • AgiBot World (real robot, AgiBot hardware) out-of-distribution scenarios
    0.67 score (0 to 1)
    average task completion scoreAuthors' own evaluation. The OXE-pretrained comparison policy scores 0.38. Together with the seen-scenario row this is the basis for the abstract's 30 percent average improvement claim.source
  • Ablation (AgiBot World, real robot) with vs without the Latent Planner
    0.12 score (0 to 1)
    gain in average task completion scoreAuthors' own ablation. This isolates the contribution of the ViLLA latent layer, which is the paper's core architectural claim.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on real AgiBot hardware. Policies pre-trained on AgiBot World reach an average task completion score of 0.77 in seen scenarios and 0.67 out of distribution, against 0.47 and 0.38 for the same policies pre-trained on Open X-Embodiment. The abstract reports over 60 percent success on complex long-horizon and dexterous tasks and a 32 percent margin over RDT.

Fine tuning it yourself

The Hugging Face card directs fine-tuning on simulation benchmarks or custom datasets to the GitHub repository. The licence is the binding constraint for product work: CC BY-NC-SA 4.0 forbids commercial use and requires share-alike on derivatives, so a fine-tuned GO-1 cannot ship in a commercial product under those terms.

Where it helps, where it does not

Strengths

  • Open weights on Hugging Face with a published architecture, which is rare for an industrial embodied foundation model
  • Latent action tokens let it absorb action-free human video, and the paper reports the latent planner is worth an average of 0.12 task completion score
  • Reported power-law scaling with data volume from 9.2k to 1M trajectories, Pearson r = 0.97, which is a rare quantified scaling claim in robotics
  • Small backbone (InternVL2.5-2B) relative to 7B-class competitors, at 30 Hz control with a 30-step chunk
  • Backed by a genuinely large, open dataset (AgiBot World, 1M-plus trajectories, 217 tasks)

Limits

  • CC BY-NC-SA 4.0 licence blocks commercial deployment and forces share-alike on any derivative, which disqualifies it for most product work
  • No VRAM, latency or GPU requirement is published anywhere this research could verify
  • Results are reported as task completion scores between 0 and 1 rather than plain success rates in the headline comparisons, which makes cross-paper comparison harder
  • Evaluations are on AgiBot's own hardware and its own dataset, so the reported margin over OXE-pretrained policies partly measures a data-distribution match
  • Superseded internally by GO-2 in April 2026

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the clientDrive a real arm now