SU

OpenVLA-OFT

Stanford University, USA · February 2025

Foundation VLAOpen weights, MITAlso written OpenVLA-OFT+, OFT recipe, Optimized Fine-Tuning
Parameters
7.5 B
The released LIBERO-Spatial checkpoint reports 7,541,237,184 BF16 parameters on Hugging Face, the same count as base OpenVLA.
GPU memory
14 GB
weights at bf16, computed. Fits a 24 GB card
Inference latency
72.9 ms
72.9 ms per 8-step action chunk on an NVIDIA A100, averaged over 100 queries with one 224 by 224 px image plus a LIBERO...
Weights
MIT
MIT (LICENSE file: Copyright 2025 Moo Jin Kim, Chelsea Finn, Percy Liang).

What it is

OpenVLA-OFT is not a new backbone but an optimized fine-tuning recipe for OpenVLA that combines parallel decoding, action chunking, continuous actions and an L1 regression objective. It replaces the 7 sequential decoder passes of base OpenVLA with a single forward pass per chunk, raising measured throughput on an A100 from 4.2 Hz to 109.7 Hz and cutting latency from 239.6 ms to 72.9 ms. On LIBERO it reaches a 97.1% four-suite average, ahead of fine-tuned pi-zero at 94.2%. The augmented OFT+ variant adds FiLM and drives a real bimanual ALOHA at 25 Hz on clothes folding and scooping. The catch is that every deployment requires its own fine-tuning run, since the released checkpoints are suite-specific rather than generalist.

Architecture

Backbone
OpenVLA (Llama 2 7B with fused DINOv2 and SigLIP encoders), adapted with LoRA; the vision tower and projector stay frozen
Action head
Parallel decoding of a whole action chunk in one forward pass, with a 2-block MLP ResNet head that regresses continuous actions under an L1 objective. A diffusion head variant is evaluated as an alternative.
Action chunk
8 steps per forward pass
Control rate
109.7 Hz
Parameters
The released LIBERO-Spatial checkpoint reports 7,541,237,184 BF16 parameters on Hugging Face, the same count as base OpenVLA. The continuous action head ships as a separate 0.30 GB file and the LoRA adapter as a separate 0.48 GB file, so neither is inside that parameter count. Chunk size is 8 for LIBERO and 25 for the ALOHA experiments.
Pretraining data
No new pretraining. OFT inherits OpenVLA's 970k Open X-Embodiment episodes and is a fine-tuning recipe on top of it. Each fine-tuning job ran 50K to 150K gradient steps at total batch size 32 on 8 A100 or H100 80 GB GPUs, taking 1 to 2 days.
Embodiments
LIBERO simulated Franka Panda, ALOHA bimanual, two ViperX 300 S arms, 25 Hz controller, 14-dimensional joint targets, WidowX (BridgeData V2, appendix scaling experiment with over 50k demonstrations)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 7.5 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3228 GBTraining master weights, rarely used for inference
bf16 or fp1614 GBThe usual way these checkpoints are served
int87.0 GBQuantised, expect some loss of precision on fine motions
int43.5 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference latency, reported72.9 ms on 72.9 ms per 8-step action chunk on an NVIDIA A100, averaged over 100 queries with one 224 by 224 px image plus a LIBERO instruction (paper Table II). With a wrist image and proprioception added the same setup measures 112.0 ms and 71.4 Hz. Base OpenVLA on the same hardware measures 239.6 ms and 4.2 Hz. source
Control rate109.7 Hz
Checkpoint download15.94 GB

The OpenVLA-OFT paper publishes no GPU memory figure. Size it from the 7.5 B parameter count and the OpenVLA fine-tuning notes instead.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO inference efficiency, 7-dimensional actions, 100 queries on an NVIDIA A100
    109.7 Hz
    action generation throughputReported by the OFT authors, Table II. Base OpenVLA measures 4.2 Hz, parallel decoding alone 15.9 Hz, parallel decoding plus chunking 108.8 Hz. With wrist image and proprioception added the figure is 71.4 Hz.source
  • LIBERO average over Spatial, Object, Goal and Long; inputs third-person image, wrist image, proprioception, language; filtered dataset; 500 trials per suite
    97.1%
    success rateReported by the OFT authors, Table I. Fine-tuned pi-zero reaches 94.2%, pi-zero plus FAST 85.5% and base OpenVLA 76.5% in the same table.source
  • LIBERO average over four suites with third-person image and language only, filtered dataset, 500 trials per suite
    95.3%
    success rateReported by the OFT authors, Table I. This is the input-matched comparison against Diffusion Policy from scratch (72.4%), fine-tuned Octo (75.1%), DiT Policy (82.4%) and base OpenVLA (76.5%).source
  • LIBERO LIBERO-Long, full OFT configuration, 500 trials
    94.5%
    success rateReported by the OFT authors, Table I. Base OpenVLA scores 53.7% on the same suite, so this is the largest single-suite gain of the recipe.source
  • LIBERO LIBERO-Long with the diffusion action head at 50 train and 50 test denoising steps
    91.1%
    success rateReported by the OFT authors, Table II. The diffusion variant beats L1 regression by 0.4 points on this suite but costs 1.907 s latency at 50 steps, and collapses to 0.0% if reduced to a single test-time step.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on a real bimanual ALOHA setup with two ViperX 300 S arms, three cameras and a 25 Hz controller (reduced from the platform's native 50 Hz) on four tasks: fold shorts (20 demos), fold shirt (30 demos), scoop X into bowl (45 demos) and put X into pot (300 demos). The abstract reports that OpenVLA-OFT+ outperforms fine-tuned pi-zero and RDT-1B as well as ACT and Diffusion Policy trained from scratch by up to 15 percentage points absolute in average success rate. The per-task scores are only given as a bar chart (Figure 4), so exact per-task values are not extractable from the paper text.

Fine tuning it yourself

LoRA on the LLM backbone with the vision tower and projector frozen, plus a separately trained MLP action head. Chunk size 8 for LIBERO and 25 for ALOHA, executed fully open loop before requerying. Datasets in the LIBERO RLDS or ALOHA HDF5 conventions; the ALOHA tasks were trained from 20 to 300 demonstrations each. Reference runs used 8 A100 or H100 80 GB GPUs for 1 to 2 days; fewer GPUs work via gradient accumulation.

Where it helps, where it does not

Strengths

  • Turns a slow autoregressive VLA into a real-time one: 26x throughput on single-arm tasks and 43x with 25-step chunks, without changing the base model
  • Highest published LIBERO average among the models in this comparison at 97.1%, beating fine-tuned pi-zero (94.2%) with a weaker and less pretrained base model
  • Adds inputs that base OpenVLA cannot take: wrist camera, second wrist camera and proprioceptive state, doubling visual patch embeddings from 256 to 512 while staying above 71 Hz
  • Documented VRAM numbers per configuration on the project website, which is rare and makes hardware planning straightforward
  • FiLM conditioning (the OFT+ variant) fixes the language-following failure that appears on bimanual tasks with visually dominant distractors

Limits

  • This is a fine-tuning recipe, not a general-purpose pretrained policy. The released checkpoints are task-suite specific (for example LIBERO-Spatial), so every deployment starts with its own fine-tuning run of 50K to 150K steps on 8 large GPUs.
  • Full action chunks are executed open loop before the model is requeried, so the policy cannot react to disturbances inside a chunk.
  • Language following on ALOHA drops to chance level when FiLM is removed, which means the plain OFT recipe is not sufficient for language-conditioned bimanual work.
  • The diffusion head variant needs its full 50 denoising steps: at 1 test-time step LIBERO-Long success falls to 0.0%, so the cheaper sampling shortcuts used by other diffusion policies do not transfer.
  • The 0.07 ms and 0.321 ms latency figures printed in Section II of the paper contradict Table II and appear to be a unit typo; use the Table II values in seconds.
  • Training minimum of 25.6 GB puts a 24 GB consumer card just out of reach for fine-tuning.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.