NAT VLA 450M

NAT VLA 450M is a robot policy built on Liquid AI’s LFM2-VL-450M. It turns images, natural-language instructions and robot state into time-ordered NAT action tokens for streaming control. LeRobot supports fine-tuning and token-prefix real-time chunking.

Papers & reports (1)

NAT: Scaling Causal Neural Action Tokenizers Across Embodiments

Paper available

Authors & affiliations (3)

Author order: Original order

  1. Ahad Jawaid

    Not provided

  2. Alexander Swerdlow

    Stanford University

  3. Yu Xiang

    The University of Texas at Dallas

First published
Not provided
Latest paper revision
Not provided
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

How to access and use

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

Code and training / fine-tuning

Access
Public access

SourceVerified on · 2026-11-22

Code license
Apache-2.0

Verified on · 2026-11-22

Model files

Model files
NAT VLA 450M PretrainedAccessWeights licenseCommercial use
ConditionalTerms

Verified on · 2026-11-22

Conditions

LFM Open License v1.0 applies to the policy checkpoint. See section 5 for the annual revenue threshold of USD 10 million and the qualified nonprofit research provision.

Release information

SourceVerified on · 2026-11-22

Checks before deployment

Fine-tune on target robot data. With the default tokenizer, streaming and token-prefix RTC require at most seven action dimensions; wider actions are decoded in groups and returned as a full chunk. The paper’s real-robot results used a separately fine-tuned Trossen AI Solo policy. 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
Greedy action-token decoding; streaming and token-prefix RTC for up to 7 dimensions
Dataset format
LeRobot dataset with action min/max and robot-state q01/q99
Runtime and frameworks
LeRobot / PyTorch / Transformers
Fine-tuning requirements
Target-robot dataset, camera configuration, robot setup and control mode
Deployment requirements
LeRobot main at e5be0c1 or later with nat_vla extra; Python ≥3.12; Transformers ≥5.22.0,<5.23.0 at e5be0c1

Architecture and inputs / outputs

Model type
Action policy
Architecture
VLA
Backbone
Liquid AI LFM2-VL-450M
Weight parameters
540222460 parameters
Action generation
Autoregressive
Input
Language / RGB images / Robot state
Output
32 steps; 14 tokens per group of up to 7 dimensions

Scope and training

Base model

Liquid AI LFM2-VL-450M

Code license

Apache-2.0

Embodiment scope

Cross embodiment

Implementation source

Official

Functional purpose

action_control

Output

Actions

Prediction output

Time-ordered NAT action tokens

Robot form

Robotic arm

Task scope

Multiple tasks

Training data scale

1,800 hours of VLA pretraining data

Training data sources

BridgeData V2 / DROID

Training

Imitation learning / Pretraining / Supervised fine-tuning

Submit a source or correction