OpenVLA-OFT
Optimized Fine-Tuning (OFT) recipe for OpenVLA from Stanford: replaces autoregressive token decoding with parallel decoding of continuous action chunks trained with an L1 regression objective (optionally FiLM for stronger language grounding, "OFT+"), improving inference speed and task success when adapting OpenVLA to new robots.
open weightsverified 2026-10-24paper →
Architecture
Compute requirements
| Memory | Precision | Conditions | |
|---|---|---|---|
| train | ≥27 GB · ≤80 GBsrc | bf16 | Upstream openvla-oft README: between 1 and 8 GPUs with 27–80 GB each, depending on the training setup, with the default bfloat16 data type. Not a LeRobot figure; batch size not stated in the README. |
| inference | ≥16 GBsrc | — | Upstream openvla-oft README: 1 GPU with ~16 GB VRAM for LIBERO sim tasks (~18 GB for ALOHA robot tasks). Precision not stated in this line; the quick-start loads without 8/4-bit quantization. |
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running OpenVLA-OFT 5 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| ALOHA (stationary) | works | T1 · upstream documented | 2026-10-24 1d ago | Real-robot results use the OFT+ variant (OFT plus FiLM language conditioning) on the original ALOHA (two ViperX 300 S arms, 3 cameras), run at 25 Hz instead of ALOHA's native 50 Hz. |
| Franka Research 3 | works with caveats | T3 · community reported | 2026-10-25 today | Community-reported only (single third-party repo, ww249/Openvla-oft-for-Franka-Research-3); not documented by the OpenVLA-OFT authors or maintainer-tested. |
| Franka Emika Panda | partial | T1 · upstream documented | 2026-10-25 today | Simulation only: LIBERO; no real-robot training or inference documented. |
| I2RT YAM | works with caveats | T3 · community reported | 2026-10-24 1d ago | Community-reported only; not documented upstream or maintainer-tested (evidence is a baseline in the MolmoAct2 paper, not the OpenVLA-OFT authors). |
| WidowX 250 S | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | The WidowX result is an appendix experiment: OpenVLA-OFT (without FiLM) trained on all of BridgeData V2 and evaluated on a subset of OpenVLA's Bridge tasks (69.2% vs. 65.8% for OpenVLA). |
Framework support
Checkpoints
- moojink/openvla-7b-oft-finetuned-libero-spatial · Used in the README quick-start (MIT tag on the Hub).
- moojink/openvla-7b-oft-finetuned-libero-10
- moojink/openvla-7b-oft-finetuned-libero-spatial-object-goal-10
Related terms
Action chunking · Action tokenization · Embodiment gap · Sim-to-real transfer · Vision-language-action model
Sources
- T1 · upstream documentedpapersource 2025-02-27verified 2026-10-24 (1d ago)
OFT integrates parallel decoding, action chunking, continuous actions and an L1 regression objective.
“an instantiation of an Optimized Fine-Tuning (OFT) recipe that integrates parallel decoding and action chunking, continuous action representations, and an L1 regression objective”
https://arxiv.org/abs/2502.19645 - T1 · upstream documentedofficial-docssource 2025-09-09verified 2026-10-24 (1d ago)
Upstream inference requirement for LIBERO sim tasks (~16 GB; ALOHA ~18 GB).
“* 1 GPU with ~16 GB VRAM for LIBERO sim benchmark tasks”
https://github.com/moojink/openvla-oft/blob/e4287e94541f459edc4feabc4e181f537cd569a8/README.md - T1 · upstream documentedofficial-docssource 2025-09-09verified 2026-10-24 (1d ago)
Upstream training requirement, 1-8 GPUs with 27-80 GB in bfloat16.
“* Between 1-8 GPUs with 27-80 GB, depending on the desired training setup (with default bfloat16 data type).”
https://github.com/moojink/openvla-oft/blob/e4287e94541f459edc4feabc4e181f537cd569a8/README.md