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

TinyVLA

Midea Group·East China Normal University·Shanghai University·Syracuse University·Beijing Innovation Center of Humanoid Robotics · 2024-09

VLAAutoregressive & latent-action VLAS 422M / B 740M / H 1.3Bdiffusion

Attaches a diffusion head to a small VLM without robot-data pretraining and trains with LoRA (5.0% of transformer parameters).

Paper

TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic ManipulationarXiv 2409.12514

Overview

Category

Autoregressive & latent-action VLA

Type

VLA

Released

2024-09

Organization

Midea Group·East China Normal University·Shanghai University·Syracuse University·Beijing Innovation Center of Humanoid Robotics

Open weights

Partial

Parameters

S 422M / B 740M / H 1.3B (trainable parameters 101M/138M/143M)

Architecture

Backbone LLM

Pythia (TinyVLA-S/B/H)

Vision encoder

LLaVA framework vision encoder

Action head

Diffusion policy head (diffusion)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Pythia LLM from the LLaVA framework (TinyVLA-S/B/H) + diffusion policy head, 7-dim continuous actions

Training

Pretraining accelerators

N/A (no robot pretraining; VLM pretraining hardware not disclosed)

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed (no GPU·time·batch·LoRA rank stated)

Data type

Real-robot teleoperation + simulation·synthetic + web vision-language (LLaVA data for the VLM)

Data scale

100 trajectories per Franka task; 50 MetaWorld demos

Main datasets

In-house Franka/UR5 teleoperation, MetaWorld, LLaVA

Tasks & evaluation

MetaWorld 50 tasks 31.6%; real Franka 5 tasks 94.0% (+25.7% over OpenVLA); bimanual UR5 44.5% (OpenVLA 0%)

Robot

Franka single arm, UR5 bimanual, MetaWorld simulation

Inference

Inference device & speed

1 × A6000 14 ms (TinyVLA-H), OpenVLA 292 ms

Inference memory

Not disclosed

Weights bf16 (calc.)

2.6 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning) only

Base control

Fixed · No base control (Franka·UR5 arms)

Notes

Key features

Attaches a diffusion head to a small VLM without robot-data pretraining and trains with LoRA (5.0% of transformer parameters). Latency on A6000 is 14 ms, about 1/20 of OpenVLA (292 ms).

Notes

A diffusion-head VLA rather than an autoregressive token model, but included since it is in the assigned scope. Only the base VLMs (about 400M/700M/1.3B) are released on HF; no policy checkpoint was found. Code released 2025-02-17. Parameters (B) based on the largest variant, TinyVLA-H.

Verification

Confirmed