VLA Gallery / All models / Autoregressive & latent-action VLA

NORA

SUTD(DeCLaRe Lab)·Lambda Labs · 2025-04

VLAAutoregressive & latent-action VLA3Bdiscrete tokens

A 3B autoregressive VLA adding FAST+ action tokens to Qwen2.5-VL-3B.

Paper

NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied TasksarXiv 2504.19854

Overview

Category

Autoregressive & latent-action VLA

Type

VLA

Released

2025-04

Organization

SUTD(DeCLaRe Lab)·Lambda Labs

Open weights

Yes

Parameters

3B

Architecture

Backbone LLM

Qwen2.5-3B · Qwen2.5-VL-3B config

Vision encoder

Qwen2.5-VL ViT

Action head

FAST+ tokens (discrete tokens)

Chunk

1 (NORA) · 5 (Long)

Control rate

Not disclosed

Architecture (original summary)

Qwen2.5-VL-3B; FAST+ tokenizer (2048 tokens added to vocabulary), chunk 1 (NORA)/5 (NORA-Long), 7-DoF actions

Training

Pretraining accelerators

8 H100 (1 node)

Pretraining duration

About 3 weeks

Pretraining steps & batch

1.1M steps, batch 256, peak lr 5e-5 after 50k-step warmup, cosine decay

GPU/TPU-hours

4,000 (Stated (approx.) · H100-hours)

Fine-tuning

150 epochs, batch 128, lr 5e-5; hardware and time not disclosed

Data type

Real-robot teleoperation

Data scale

970k demos

Main datasets

Open X-Embodiment

Tasks & evaluation

Real WidowX-250 9 tasks 56.7%; LIBERO 87.9% (NORA-Long)

Robot

WidowX-250, LIBERO simulation

Inference

Inference device & speed

GPU memory about 8.3 GB; latency not disclosed

Inference memory

About 8.3 GB

Weights bf16 (calc.)

6 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning) only

Base control

Fixed · No base control (WidowX arm·LIBERO)

Notes

Key features

A 3B autoregressive VLA adding FAST+ action tokens to Qwen2.5-VL-3B. Pretrained for about 3 weeks on one node of 8 H100.

Notes

Verification added Lambda Labs to affiliations. A follow-up NORA-1.5 exists.

Verification

Corrections applied