VLA GalleryAll modelsGallery

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

OpenVLA-OFT

Stanford · 2025-02

VLAAutoregressive & latent-action VLAOpenVLA 7B + additional 279Mregression

Changes the OpenVLA fine-tuning recipe to parallel decoding·action chunks·L1 regression, raising throughput 26~43×.

View architecture diagram in the galleryPaper ↗

Paper

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and SuccessarXiv 2502.19645

Overview

Category
Autoregressive & latent-action VLA
Type
VLA
Released
2025-02
Organization
Stanford
Open weights
Yes
Parameters
OpenVLA 7B + additional 279M (LIBERO: LoRA 111M·head 151M·proprio 17M)/853M (ALOHA OFT+, including FiLM 456M)

Architecture

Backbone LLM
Llama 2 7B · OpenVLA
Vision encoder
SigLIP + DINOv2 (600M)
Action head
Parallel decoding + L1 regression MLP (regression)
Chunk
8 (LIBERO) · 25 (ALOHA)
Control rate
25 Hz (ALOHA)
Architecture (original summary)
OpenVLA-7B; parallel decoding + continuous action L1 regression (4-layer MLP head), chunk 8 (LIBERO)/25 (ALOHA), OFT+ adds FiLM

Training

Pretraining accelerators
N/A (uses the OpenVLA checkpoint)
Pretraining duration
N/A
Pretraining steps & batch
N/A
GPU/TPU-hours
Not disclosed
Fine-tuning
8 × A100/H100 80GB, 50–150K steps, batch 64–128 (LIBERO)/32 total (ALOHA, 4 per GPU), 1–2 days per run; minimum 25.6 GB (LIBERO 1 image, batch 1) · 38.6 GB (ALOHA 3 images + FiLM), recommended 44.1 GB (LIBERO 1 image, batch 8) · 62.5 GB (2 images + proprio, batch 8) · 73.5 GB (ALOHA, batch 4); FAQ recommends 4 or 8 GPUs
Data type
Simulation·synthetic + real-robot teleoperation
Data scale
About 500 demos per LIBERO suite; ALOHA 20/30/45/300 demos per task
Main datasets
LIBERO, in-house ALOHA teleoperation
Tasks & evaluation
LIBERO average 97.1% (97.6/98.4/97.9/94.5); 4 real ALOHA tasks (fold shorts, fold shirt, scoop X into bowl, put X into pot)
Robot
LIBERO simulation single arm, ALOHA bimanual (25 Hz control)

Inference

Inference device & speed
A100 109.7 Hz · 0.0729 s (single image), 71.4 Hz with extra inputs; VRAM about 16 GB (LIBERO)/18 GB (ALOHA)
Inference memory
About 16 GB (LIBERO) · 18 GB (ALOHA)
Weights bf16 (calc.)
14 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Imitation-learning fine-tuning (L1 regression)
Base control
Fixed · No base control (LIBERO single arm·ALOHA fixed bimanual)

Notes

Key features
Changes the OpenVLA fine-tuning recipe to parallel decoding·action chunks·L1 regression, raising throughput 26~43×. Raises the LIBERO average from 76.5% to 97.1%.
Notes
Included as OpenVLA's standard follow-up recipe even though it does not use tokens at inference. Parameters (B) are based on OpenVLA 7B, excluding the additional 279M/853M. The README warns to evaluate on the same GPU used for training.
Verification
Confirmed

Other models in this category