WorldFoundry
English中文

OpenVLA-OFT

OpenVLA-OFT: `image`, `prompt`, and `state` in; `action_trace` and `openvla-oft_action_trace.json` out; 5 variants; dedicated `worldfoundry-vla-openvla-compat-cu121`.

Edit this pageGitHubMarkdown

OpenVLA-OFT applies optimized fine-tuning to OpenVLA — parallel decoding, continuous action heads, and chunking — and is fully GPU-validated in this catalog: an A100 restored all 7.54B parameters, produced deterministic finite 8x7 action chunks across multiple inputs at 14.47 GiB peak VRAM, and all five declared LIBERO checkpoints completed real inference (MIT license).

Official paper teaser: OpenVLA-OFT fine-tunes OpenVLA with parallel decoding for faster robot control.
Official paper teaser: OpenVLA-OFT fine-tunes OpenVLA with parallel decoding for faster robot control.

The paper is arXiv 2502.19645. Project page: openvla-oft.github.io. Code: moojink/openvla-oft.

Catalog aliases include openvla_oft and openvla-oft-libero.

Build your run

Pipeline
OpenVLAOFTPipelineworldfoundry.pipelines.component_pipelines
TaskDefault default
Environmentworldfoundry-vla-openvla-compat-cu121
DeviceCUDA 12.1
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  openvla-oft-libero-spatial \
  --pipeline.task-profile default \
  --json

Variants

  • openvla-oft-libero-spatial
    StatusVerified

    vla.policy_rollout

    Runtime profileopenvla-oft

    Pipeline bindingopenvla-oft

  • openvla-oft-libero-object
    StatusVerified

    vla.policy_rollout

    Runtime profileopenvla-oft

    Pipeline bindingopenvla-oft

  • openvla-oft-libero-goal
    StatusVerified

    vla.policy_rollout

    Runtime profileopenvla-oft

    Pipeline bindingopenvla-oft

  • openvla-oft-libero-10
    StatusVerified

    vla.policy_rollout

    Runtime profileopenvla-oft

    Pipeline bindingopenvla-oft

  • openvla-oft-libero-spatial-object-goal-10
    StatusVerified

    vla.policy_rollout

    Runtime profileopenvla-oft

    Pipeline bindingopenvla-oft

What it is for

Use the OFT LIBERO specialists when you want parallel-decoded 8×7 action chunks on the OpenVLA 7B backbone — not the base OpenVLA card.

Launch worldfoundry-eval run openvla-oft-libero-spatial. Required: image, prompt, state. Artifacts: action_trace and openvla-oft_action_trace.json. Swap openvla-oft-libero-object, -goal, -10, or the combined spatial-object-goal-10 card for the other splits.

Related suite: LIBERO.

Runtime contract

Required: image, prompt, state. Cameras: full_image and wrist_image. Proprio is 8-D; action chunk is 8×7. Official demo observation: experiments/robot/libero/sample_libero_spatial_observation.pkl. Status: Verified — all five declared LIBERO checkpoints completed A100 inference. Upstream memory hint is about 16 GB for LIBERO.

Running in WorldFoundry

Dedicated environment worldfoundry-vla-openvla-compat-cu121 (Python 3.10, CUDA 12.1). Stage the matching moojink/openvla-7b-oft-finetuned-libero-* snapshot — the runtime does not download weights. Launch worldfoundry-eval run openvla-oft-libero-spatial. License: MIT.