SANA-Robot-AR RoboTwin 2.0

Papers and technical reports

SANA-Robot-AR RoboTwin 2.0 is a 4.5B-parameter world action model post-trained on 2,500 clean-scene demonstrations across 50 simulated tasks. It uses three camera views, robot state and a language instruction to predict future video and 32 steps of end-effector control for the dual-arm ALOHA-AgileX.

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 data
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

Training data

Some training datasets are public; this does not establish access to the complete training mixture.

Model files

Model files
SANA-Robot-AR RoboTwin 2.0 15kAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Release information

SourceVerified on · 2026-11-22

Checks before deployment

The published results cover the simulated ALOHA-AgileX in RoboTwin 2.0. Other robots, camera arrangements, tasks or scenes require 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
Diffusion
Input
Language / RGB images / Robot state
Output
Dual-arm end-effector pose control, 32 steps per chunk
Released
November 22, 2026

Scope and training

Dual-arm end-effector pose control, 32 steps per chunk

horizon 32

Output

Actions / Video

Prediction output

Action chunk and future video at a frame stride of 4

Lineage

Fine-tuned from: SANA-Robot · SANA-Robot-AR ↗

Training data

  • RoboTwin 2.0

    Fine-tuningAvailable

    2500 episodes · MIT License

Evaluation results

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

Full parameters and verification notes (model card)

The PyTorch file contains policy parameters under state_dict. Use the supplied metadata and normalization files. Its training metadata specifies 50 Hz. A head camera (384 × 480) and two wrist cameras (192 × 256 each) use a T-shaped layout. Gemma 2 2B IT (Google terms apply) and the Diffusers-layout LTX-2.3 video VAE are separate downloads. The model card points to eval/robotwin for the policy server and adapter and configs/robotwin_ar.yaml for fine-tuning.

Submit a source or correction