DynamicVLA LIBERO

Papers and technical reports

S-Lab at Nanyang Technological University initialized this DynamicVLA release from its DOM model and fine-tuned it in LeRobot on four LIBERO simulation suites. It runs 10 action steps from each prediction and provides downloadable Safetensors weights.

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

Public code resources
Public access
Model files
Public access
Training data
Public access

Code and training / fine-tuning

Access
Public access

SourceVerified on · 2026-11-22

Code license
S-Lab License 1.0

Verified 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
DynamicVLA LIBERO · SafetensorsAccessWeights licenseCommercial use
ProhibitedTerms

Verified on · 2026-11-22

Conditions

S-Lab License 1.0 permits non-commercial use and redistribution under its conditions. Contact the contributors for commercial terms.

Release information

SourceVerified on · 2026-11-22

Checks before deployment

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

Inference mode
Action chunking, streaming disabled
Runtime and frameworks
LeRobot
Fine-tuning requirements
40,000 steps, global batch size 72 (3 GPUs × 24), AdamW, 1,000 warmup steps, cosine learning rate from 0.0001 to 0.0000025
Deployment requirements
LeRobot with the LIBERO environment (MuJoCo)

Architecture and inputs / outputs

Model type
Action policy
Architecture
VLA
Backbone
SmolLM2-360M
Action generation
Flow matching
Input
Language / RGB images / Robot state
Output
7-dimensional LIBERO action
Released
November 22, 2026

Scope and training

Output

Actions

Prediction output

20 predicted action steps, 10 executed per prediction

Task scope

Multiple tasks

Training data sources

HuggingFaceVLA LIBERO

Training

Supervised fine-tuning

Training data

Evaluation results

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

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Overall success rate: 61.4 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

LIBERO-Spatial: 78 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

LIBERO-Object: 71.9 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

LIBERO-Goal: 42.1 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

LIBERO-10: 54.9 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Camera: 38.5 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Robot: 63.7 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Language: 75.5 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Lighting: 87.2 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Background: 52.9 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Noise: 46.2 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Publisher evaluation · Reported; not independently reproduced

LIBERO-Plus · DynamicVLA zero-shot ↗

Object layout: 71.7 %

4976dc3

LeRobot libero_plus, 256 × 256 rendering, hard resets and batch size 1. The simulator’s fog noise uses np.float64 for NumPy 2 compatibility.

Author-reported zero-shot evaluation of all 10,030 registered variants, one episode per variant. The checkpoint was fine-tuned on LIBERO only, with no prior exposure to LIBERO-Plus data.

Submit a source or correction