SANA-Robot-AR

Papers and technical reports

SANA-Robot-AR is a 4.5B-parameter world action model that predicts 32-step action chunks and future video. It combines windowed attention with Gated DeltaNet memory. The public pretrained weights are intended for fine-tuning on a target robot or benchmark.

View full model description

Initialization for further fine-tuning on target robot demonstrations or simulation benchmarks.

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 PretrainAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Release information

SourceVerified on · 2026-11-22

Checks before deployment

Direct robot control with the pretrained weights has not been validated. Load the text encoder and video VAE separately. 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
Fine-tuning requirements
Target robot demonstrations, Gemma 2 2B IT and LTX-2.3 VAE

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
32 control steps per action chunk
Released
November 22, 2026

Scope and training

Output

Actions / Video

Prediction output

Action chunk and future video at a frame stride of 4

Training

Pretraining

Derived models

SANA-Robot-AR RoboTwin 2.0 ↗Fine-tuned model

SANA-Robot-AR LIBERO ↗Fine-tuned model

Full parameters and verification notes (model card)

Each chunk predicts 32 control steps and encodes video at a frame stride of 4. Softmax attention reads the first frame and the previous chunk, while Gated DeltaNet summarizes preceding chunks in a fixed-size state.

External dependencies: Gemma 2 2B IT (Google terms acceptance) and LTX-2.3 video VAE (Diffusers layout).

Fine-tuning recipes: RoboDojo, RoboTwin 2.0 and LIBERO. Each has a separately released fine-tuned checkpoint, and the benchmark scores belong to those variants.

Submit a source or correction