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

VLA-Adapter

BUPT·Westlake University·Zhejiang University·HKUST(GZ)·OpenHelix Team · 2025-09

VLAAutoregressive & latent-action VLAQwen2.5-0.5B backbone + 197.2M trainable parametersregression

Reached LIBERO 97.3% with a 0.5B backbone and Bridge Attention policy without robot pretraining.

Paper

VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action ModelarXiv 2509.09372

Overview

Category

Autoregressive & latent-action VLA

Type

VLA

Released

2025-09

Organization

BUPT·Westlake University·Zhejiang University·HKUST(GZ)·OpenHelix Team

Open weights

Yes

Parameters

Qwen2.5-0.5B backbone + 197.2M trainable parameters (policy 97.3M); total not stated

Architecture

Backbone LLM

Qwen2.5-0.5B

Vision encoder

SigLIP + DINOv2 (Prismatic)

Action head

Bridge Attention policy (L1 regression) (regression)

Chunk

8 steps

Control rate

Not disclosed

Architecture (original summary)

Prismatic VLM (Qwen2.5-0.5B, same base as MiniVLA); L1 regression with Bridge Attention policy, chunk H=8, no action tokens

Training

Pretraining accelerators

N/A (no robot pretraining)

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Paper: 4× H100, batch 16·up to 150,000 steps (README per task: LIBERO-Spatial about 5 h, Object under 1 h, Goal 3 h, Long half a day, CALVIN 8 h); 8 h on 1× 32GB consumer GPU (LIBERO-Object 99.2%, batch 8·LoRA 64); minimum 9.6 GB (batch 1·LoRA 64)

Data type

Simulation·synthetic

Data scale

LIBERO·CALVIN demos (count not stated)

Main datasets

LIBERO, CALVIN

Tasks & evaluation

LIBERO average 97.3%; CALVIN ABC→D average length 4.42

Robot

LIBERO·CALVIN simulation

Inference

Inference device & speed

219.2 Hz, 0.0365 s (OpenVLA-OFT 71.4 Hz); GPU not stated

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning) only

Base control

Fixed · No base control (LIBERO·CALVIN arm sim)

Notes

Key features

Reached LIBERO 97.3% with a 0.5B backbone and Bridge Attention policy without robot pretraining. Inference runs at 219.2 Hz, faster than OpenVLA-OFT (71.4 Hz).

Notes

A regression-head VLA rather than token-based, included as a reference row for the 2025 compute-efficiency comparison. Verification resolved the attribution of 4×H100 (paper) and '8 hours' (one 32GB consumer GPU, LIBERO-Object).

Verification

Confirmed