Open Source RoboticsRobots, software & learning

OpenVLA-OFT

Robotics model

Optimized Fine-Tuning (OFT) recipe and five MIT-licensed 7.54B-parameter LIBERO checkpoints built on OpenVLA-7B, replacing autoregressive token decoding with parallel continuous L1-regression action chunking, multi-camera inputs, proprioception, and optional FiLM language conditioning. [1] [2] [3] [4]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
VLA researchers benchmarking on LIBERO and ALOHA; Robotics practitioners fine-tuning OpenVLA-7B with LoRA on custom RLDS datasets [1] [3] [5]
Parameters
7,541,237,184 (~7.54B BF16 backbone params per HF safetensors metadata, plus lightweight MLP action head and… [4]
Embodiments
Franka Panda (LIBERO simulation), ALOHA bimanual manipulator (real robot, 14-DoF), WidowX (Bridge constants defined… [3] [5] [6]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

OpenVLA-OFT (moojink/openvla-oft, arXiv:2502.19645) is an MIT-licensed fine-tuning recipe and checkpoint family built on openvla/openvla-7b. Instead of decoding discretized action tokens autoregressively one step at a time, OpenVLA-OFT adapts the 7.54B-parameter Prismatic VLM backbone with LoRA (r=32), parallel action-chunk decoding, a continuous MLP L1-regression head (or optional diffusion head), a proprioceptive state projector, multi-camera RGB inputs, and optional FiLM language conditioning (OFT+). [1] [3] [4] [5]

Model card summary

Architecture

OpenVLA-7B Prismatic VLM backbone (DINOv2 + SigLIP vision encoders and Llama-2 7B) adapted with parallel decoding, continuous MLP L1-regression action head (or optional diffusion head), proprioception projector, multi-image inputs, and optional FiLM language-vision modulation (OFT+).

[1] [3] [5]

Base model

openvla/openvla-7b

[3] [4]

Parameter count

7,541,237,184 (~7.54B BF16 backbone params per HF safetensors metadata, plus lightweight MLP action head and proprioception projector .pt modules)

[4]

Input modalities
rgbproprioceptionlanguage

2 RGB views (third-person + wrist) and 8-D proprioception for LIBERO; 3 RGB views (third-person + left/right wrist) and 14-D proprioception for ALOHA.

[1] [3] [5]

Output modalities
ee-pose-deltajoint-positionsgripper-command

7-DoF relative EEF pose + gripper for LIBERO and Bridge; 14-DoF absolute bimanual joint angles for ALOHA.

[3] [5] [6]

Action space

Continuous parallel-decoded action chunks: 8-step x 7-DoF relative EEF deltas (BOUNDS_Q99) for LIBERO, 25-step x 14-DoF absolute joint angles (BOUNDS) for ALOHA, and 5-step x 7-DoF for Bridge per prismatic/vla/constants.py.

[5] [6]

Control frequency

25 Hz documented for ALOHA setup (with 25-step action chunks; 50 Hz supported with 50-step chunks); LIBERO closed-loop rate not pinned as a single Hz value in repo docs

[5]

Trained / supported embodiments
  • Franka Panda (LIBERO simulation)
  • ALOHA bimanual manipulator (real robot, 14-DoF)
  • WidowX (Bridge constants defined in codebase)

[3] [5] [6]

Training datasets
  • openvla/modified_libero_rlds (LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10 no-noops RLDS, ~10 GB)
  • Custom ALOHA bimanual demonstration datasets converted to RLDS (user-collected; not hosted as prebuilt checkpoints)

[3] [5]

Interfaces
  • Python evaluation API (get_vla, get_action_head, get_proprio_projector, get_vla_action)
  • FastAPI / uvicorn HTTP policy server (vla-scripts/deploy.py)
  • torchrun CLI fine-tuning (vla-scripts/finetune.py)
  • RLDS dataset loader

[1] [3] [5]

Checkpoints

CheckpointFormatSizeAccessPinned revision
moojink/openvla-7b-oft-finetuned-libero-spatialsafetensors14.84 GBopen6d0231af0e48
moojink/openvla-7b-oft-finetuned-libero-objectsafetensors14.84 GBopen4c89574e1c53
moojink/openvla-7b-oft-finetuned-libero-goalsafetensors14.84 GBopenc2d0f9fbbd82
moojink/openvla-7b-oft-finetuned-libero-10safetensors14.84 GBopen95220f9a3421
moojink/openvla-7b-oft-finetuned-libero-spatial-object-goal-10safetensors14.84 GBopen638918f3d1c2

Each checkpoint repo contains BF16 safetensors backbone shards plus .pt files for the MLP action head and proprioception projector (14.84 GB total per repo).

[3] [4]

Is this for you?

Use cases
  • Evaluating or fine-tuning continuous action-chunking VLA policies on LIBERO suites
  • LoRA (r=32) fine-tuning of OpenVLA-7B with multi-camera RGB, proprioception, and FiLM conditioning on bimanual ALOHA tasks
  • Serving fine-tuned VLA policies over FastAPI/uvicorn to a lightweight robot client

[1] [3] [5]

Known limitations
  • Upstream recommends evaluating on the same GPU model used for training because cross-GPU evaluation can cause substantial performance drops; merging LoRA weights on the target device may mitigate the drop
  • Paper results depend on a custom transformers v4.40.1 fork (moojink/transformers-openvla-oft), Python 3.10.14, and PyTorch 2.2.0
  • Published Hugging Face checkpoints cover LIBERO simulation suites only; real-world ALOHA checkpoints are not hosted on Hugging Face
  • Checkpoints must be evaluated with center_crop=True to match the 90%-area random crop training augmentation

[3] [5]

Safety and limitations
  • Evaluating a fine-tuned checkpoint on a different GPU architecture than the training GPU can substantially degrade performance unless LoRA adapters are merged on the target device.
  • Base Llama-2 7B weights underlying OpenVLA-7B are subject to upstream OpenVLA / Llama-2 provenance even though the OpenVLA-OFT code and released checkpoints declare MIT.
  • No safety interlocks or collision checks are included in the policy server or robot evaluation scripts.

[3] [5]

Compute & fine-tuning

Inference compute
  • Minimum VRAM: 16 GB
  • Documented hardware: NVIDIA A100 (upstream LIBERO paper evaluation)
  • Precision: bf16
  • Edge / consumer capable: No

16 GB VRAM for LIBERO (18 GB for ALOHA) in bfloat16; exact per-step latency_ms not pinned in repo README.

[1] [3]

Fine-tuning
  • Supported: Yes
  • Methods: LoRA (r=32) fine-tuning with parallel decoding and L1 regression via vla-scripts/finetune.py; Optional diffusion action head (--use_diffusion True); Optional FiLM language conditioning (--use_film True, OFT+)
  • Minimum VRAM: 25 GB
  • Recipe: Upstream fine-tuning documentation

LIBERO.md reports ~25 GB VRAM at batch size 1 per GPU and ~62 GB at batch size 8 per GPU in bfloat16; paper reproduction uses 8 GPUs for 150K steps (50K for LIBERO-Goal).

[1] [3] [5]

Host computer

Inference requires 1 GPU with ~16 GB VRAM for LIBERO and ~18 GB VRAM for ALOHA; fine-tuning requires 1–8 GPUs with 25–80 GB VRAM in bfloat16 (~25 GB at batch size 1 per GPU, ~62 GB at batch size 8 per GPU on LIBERO).

[1] [3]

Operating system
  • Linux

Standard Linux conda, CUDA, flash-attn, and ROS/ALOHA client workflow documented.

[1] [5] · Editorial assessment

Dependencies
  • Python 3.10 (3.10.14 tested)
  • PyTorch 2.2.0
  • moojink/transformers-openvla-oft (custom transformers v4.40.1 fork)
  • flash-attn==2.5.5 (for training)
  • openvla/openvla-7b (base weights for fine-tuning)
  • uvicorn / fastapi / json-numpy (for ALOHA server deployment)

[1] [3] [5]

Skills
  • Python 3.10 and conda
  • PyTorch / LoRA fine-tuning and RLDS dataset conversion
  • FastAPI / uvicorn server-client deployment (for ALOHA)

Inferred from SETUP.md, LIBERO.md, and ALOHA.md workflows.

[1] [5] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] · Model card / documentation [1]

  • Create a Python 3.10 conda environment and install PyTorch
  • Clone moojink/openvla-oft and run pip install -e . (which installs the custom transformers-openvla-oft fork)
  • Install flash-attn==2.5.5 for training and LIBERO or ALOHA requirements for evaluation
  • Load a moojink/openvla-7b-oft-finetuned-libero-* checkpoint via get_vla, get_action_head, and get_proprio_projector, or run vla-scripts/finetune.py from openvla/openvla-7b

[1] [3]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LIBERO 4-suite average (Spatial, Object, Goal, 10)97.1% average (4 task-specific checkpoints); 96.8% average (single combined 4-suite checkpoint)Upstream-reported results on NVIDIA A100 averaged over 3 random seeds (500 trials/suite, custom transformers v4.40.1 fork); not independently reproduced by OSR.

[3]

Project notes

Action chunking, multi-view inputs, and platform constants

In prismatic/vla/constants.py, the codebase defines three platform configurations: LIBERO uses 8-step action chunks over 7-DoF relative end-effector deltas (BOUNDS_Q99 normalization) with two RGB views (third-person plus wrist) and 8-D proprioception; ALOHA uses 25-step chunks over 14-DoF absolute bimanual joint angles (BOUNDS normalization, avoiding quantile clipping) at 25 Hz with three RGB views and 14-D proprioception; and Bridge defines 5-step 7-DoF chunks. Five ungated checkpoints are published on Hugging Face—four per-suite policies (libero-spatial, libero-object, libero-goal, libero-10) and one combined four-suite policy (libero-spatial-object-goal-10)—each totaling 14.84 GB across BF16 safetensors backbone shards (7,541,237,184 parameters) plus .pt weights for the MLP action head and proprioception projector. Real-world ALOHA checkpoints are not hosted on Hugging Face and must be fine-tuned from user demonstrations. [3] [4] [5] [6]

Compute requirements, deployment caveats, and upstream benchmarks

Upstream documents ~16 GB VRAM for LIBERO inference and ~18 GB VRAM for ALOHA inference in bfloat16, while LoRA fine-tuning requires ~25 GB VRAM at batch size 1 per GPU (~62 GB at batch size 8 per GPU on LIBERO). For LIBERO paper reproduction, upstream documents 8 GPUs for 150K steps (or the 50K-step checkpoint for LIBERO-Goal); for the documented real-world ALOHA put X into pot example, upstream instead specifies 8 GPUs and the 100K-step checkpoint. For real-robot ALOHA deployment, vla-scripts/deploy.py serves actions over FastAPI/uvicorn to a separate lightweight robot client environment. Three operational caveats are explicit in the upstream guides: evaluation requires --center_crop True to match training augmentations, LIBERO paper reproduction relies on a custom transformers v4.40.1 fork (moojink/transformers-openvla-oft) with Python 3.10.14 and PyTorch 2.2.0, and upstream recommends evaluating on the same GPU model used for training because cross-GPU evaluation can substantially degrade performance—merging LoRA weights into the base model on the target device via merge_lora_weights_and_save.py is noted as a possible mitigation rather than a guaranteed remedy. Upstream reports 97.1% average LIBERO success across the four task-specific checkpoints and 96.8% for the combined checkpoint (500 trials per suite, averaged over three seeds on NVIDIA A100). All benchmark figures are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [3] [5]

What we checked

None. Prepared from public upstream sources and Hugging Face API metadata without hands-on training, fine-tuning, or inference.

Updates

Upstream last updated
Sep 9, 2025 [1]
Latest release
Jun 17, 2025 [4]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
moojink/openvla-oft@e4287e94541f459edc4feabc4e181f537cd569a8; HF moojink/openvla-7b-oft-finetuned-libero-spatial@6d0231af0e48c5985f1ff86908f4674b84bc049b · combined hardware and software repository

Separate license matrix

MIT-licensed fine-tuning recipe and five ungated MIT-licensed 7.54B LIBERO VLA checkpoints built on OpenVLA-7B with public RLDS training datasets and reproducible LIBERO/ALOHA recipes.

LayerAvailability / license
Codeavailable · MIT
Verified MIT License in root LICENSE file. [1] [2]
Model weightsavailable · MIT
All 5 Hugging Face LIBERO checkpoints declare license: mit in model card YAML frontmatter. [4]
Training datasetsavailable · MIT (LIBERO upstream benchmark; modified_libero_rlds hosted publicly on Hugging Face)
LIBERO fine-tuning RLDS datasets (~10 GB) are publicly available; real-world ALOHA demonstrations are user-collected. [3]
Documentationavailable · MIT
[1] [2]

Licensing and dataset availability

Both the moojink/openvla-oft codebase (verified in the root LICENSE file) and all five Hugging Face LIBERO checkpoints declare the MIT License. The LIBERO fine-tuning datasets (openvla/modified_libero_rlds, ~10 GB in RLDS format) are publicly hosted on Hugging Face, while ALOHA fine-tuning uses user-collected demonstrations converted via rlds_dataset_builder. This review verifies OpenVLA-OFT's MIT declarations for its code and LIBERO checkpoints, but does not resolve whether underlying openvla/openvla-7b (Llama-2 7B) base-model license obligations also apply; obtain and review the relevant OpenVLA and Llama-2 license texts before finalizing commercial compliance. [2] [3] [4] [5]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile prepared from moojink/openvla-oft@e4287e94541f and five HF LIBERO checkpoints.