SANA-Robot-AR RoboDojo

Papers and technical reports

SANA-Robot-AR RoboDojo is a 4.5B-parameter world action model post-trained for the dual-arm ARX-X5 in RoboDojo. From multi-view images, robot state and language, it predicts 32-step action chunks alongside future video. The public 20,000-step EMA checkpoint supports joint and end-effector control.

View full model description

Papers and technical reports

Paper / report unconfirmed

Official introductory materials are recorded; no version-specific paper or technical report has been verified.

Last checked: 2026-11-22

How to access and use

Model files
Public access

Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes

Training / fine-tuning code

SourceVerified on · 2026-11-22

Model files

Model files
SANA-Robot-AR RoboDojo 20k EMAAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Release information

SourceVerified on · 2026-11-22

Checks before deployment

Evaluated in RoboDojo simulation with the dual-arm ARX-X5. Transfer to other robots, camera arrangements or tasks requires fine-tuning. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Runtime and frameworks
PyTorch

Architecture and inputs / outputs

Model type
Action policy
Architecture
WAM · Autoregressive
Backbone
Hybrid Diffusion Transformer (24 Gated DeltaNet layers and 8 softmax layers)
Weight parameters
4500000000
Action generation
Flow matching
Input
Language / RGB images / Robot state
Output
Joint and end-effector control, 32 steps per chunk
Released
November 22, 2026

Scope and training

Joint control for dual-arm ARX-X5.

horizon 32

End-effector pose control for dual-arm ARX-X5.

horizon 32

Output

Actions / Video

Prediction output

Action chunk and future video at a frame stride of 4

Training

Imitation learning

  • Output · Prediction output · Training · Joint control for dual-arm ARX-X5. · continuous · End-effector pose control for dual-arm ARX-X5. · continuous: SANA-Robot-AR RoboDojo model card

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Full parameters and verification notes (model card)

The 20,000-step EMA checkpoint (decay 0.999) contains policy weights in state_dict_ema. Use the supplied metadata and normalization. Training metadata specifies 25 Hz and three cameras. Its Gemma 2 2B IT text encoder and LTX-2.3 video VAE are obtained separately. Google requires accepting the Gemma terms. The model card points to eval/robodojo for the XPolicyLab adapter and to configs/robodojo_ar.yaml for the fine-tuning recipe.

Submit a source or correction