What Is OpenVLA-OFT?

OpenVLA-OFT is introduced in Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success (arXiv 2502.19645) by Moo Jin Kim, Chelsea Finn and Percy Liang, accepted to Robotics: Science and Systems (RSS) 2025, with an official project page and GitHub repository. The authors start from a practical problem: VLAs struggle with novel robot setups and need fine-tuning, yet how to fine-tune them most effectively is unclear given many possible strategies. Using OpenVLA as the representative base model, they study three design choices: how actions are decoded, how actions are represented, and which learning objective is used. OFT is the recipe their analysis informs, and OpenVLA-OFT is an instantiation of it. The paper says the recipe improves inference efficiency, policy performance and flexibility in the model’s input-output specifications. This page is a definition, not a guide to running the code.

OpenVLA-OFT is third-party research. Firsthand is not affiliated with its authors or their institution, and makes no claim that OpenVLA-OFT or any other named model was trained, fine-tuned or evaluated on Firsthand data. The site states custom egocentric collection to a written spec, fast and on demand, delivering first-person human video with 3D hand pose at 21 joints per hand and a per-joint visibility flag, calibrated stereo depth and 6-DoF IMUs at 200 Hz. The only language data tied to Firsthand capture is one task instruction per episode. Firsthand captures first-person human video, not robot or teleoperation data.

Per the paper, original OpenVLA training predicts 7 discrete action tokens per timestep one after another (autoregressive decoding), with each action dimension normalised and split into 256 bins and a next-token cross-entropy loss. The authors say this makes action chunking impractical, because a single action takes 0.33 seconds on an NVIDIA A100 GPU.

OFT changes three things. Parallel decoding feeds the decoder empty action embeddings and swaps the causal attention mask for a bidirectional one, so all actions come out in a single forward pass. Action chunking then predicts several future timesteps in that same pass. Continuous actions replace the discrete tokens: an MLP action head maps the decoder’s final hidden states directly to action values, trained with an L1 regression loss. The paper compares this with a diffusion objective and reports comparable performance from L1 with faster training convergence and inference.

Original OpenVLA processes a single camera view. The paper describes a more flexible pipeline: several camera views, each encoded into 256 patch embeddings, plus low-dimensional robot state such as joint angles and gripper state, all concatenated with the language tokens before the decoder. Adding a wrist camera image and robot state doubles the visual patch embeddings passed to the decoder from 256 to 512.

OFT+ is the augmented recipe used on the ALOHA robot. It adds feature-wise linear modulation (FiLM), which infuses the language embedding into the visual features. The authors say policies on that setup could latch onto spurious visual correlations instead of the instruction. In their ablation, removing FiLM dropped language following to 33%, equal to choosing the target at random.

On the LIBERO simulation benchmark, the abstract reports that OpenVLA-OFT raised OpenVLA’s average success rate across four task suites from 76.5% to 97.1% while increasing action generation throughput by 26 times, and describes this as a new state of the art. On a real bimanual ALOHA robot, the authors report OpenVLA-OFT+ running dexterous tasks at 25 Hz and outperforming other VLAs fine-tuned with their default recipes, named in the paper as π₀ and RDT-1B, and policies trained from scratch, by up to 15% absolute in average success rate.

These are the authors’ own evaluations on their chosen tasks. The paper does not claim that the recipe is universally better than diffusion; its limitations section says L1 regression may struggle with truly multimodal demonstrations.

No, they answer different questions. LoRA is about which weights are trained. OFT is about how actions are decoded, represented and optimised. The paper states that its study adapts OpenVLA through LoRA because the training sets are small, 500 demonstrations versus 1M for pretraining, and the project page says that because it uses LoRA there is no need for model sharding. OFT is therefore a recipe used on top of LoRA in that work, and LoRA fine-tuning for VLA models is defined on its own page.

The GitHub repository is listed with an MIT licence. Its README states that inference needs about 16 GB of VRAM for LIBERO and about 18 GB for ALOHA tasks, and that training needs between 1 and 8 GPUs with 27 to 80 GB, depending on the setup. The project page says the authors’ own jobs used 8 A100 or H100 GPUs with 80 GB, ran 50K to 150K gradient steps, and took 1 to 2 days. Read the licence text itself before a commercial program relies on it.

The paper names three open questions. Its demonstrations use a consistent strategy per task, so how OFT handles multimodal demonstrations is untested. It studies fine-tuning only, so whether the benefits carry over to pretraining is open. And language grounding was inconsistent: OpenVLA without FiLM grounded language poorly on ALOHA despite no such issue on LIBERO, and the cause is described as unclear.

OpenVLA-OFT is fine-tuned on robot demonstrations: camera images, robot state, a language instruction and robot actions. Firsthand captures first-person human video, not robot or teleoperation data, so it does not deliver the robot action trajectories or proprioception those recipes consume. Whether human egocentric data would help a policy like OpenVLA-OFT is Not stated in the sources read, and this page draws no conclusion about it.

What the site does state is custom egocentric collection, fast and on demand, to a written spec. The only language data is one task instruction per episode. Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour, and delivery is in RLDS, LeRobot, WebDataset or HDF5.

What is OpenVLA-OFT?

OpenVLA fine-tuned with the Optimized Fine-Tuning recipe from Kim, Finn and Liang. The recipe combines parallel decoding, action chunking, a continuous action representation and an L1 regression objective.

What does OFT stand for?

Optimized Fine-Tuning. The paper uses OFT for the recipe and OpenVLA-OFT for its instantiation on OpenVLA. OFT+ adds FiLM for stronger language grounding on the ALOHA robot.

Is OpenVLA-OFT the same as LoRA?

No. LoRA decides which weights are trained, while OFT changes how actions are decoded, represented and optimised. The paper adapts OpenVLA through LoRA in its study because its training sets are small.

How much faster is it than base OpenVLA?

The abstract reports 26 times higher action generation throughput on LIBERO. The paper reports 43 times higher throughput than base OpenVLA with 25-timestep chunks on ALOHA. Both are the authors’ measurements on their hardware.

Is OpenVLA-OFT open source, and what is the licence?

The official GitHub repository is listed with an MIT licence, and the abstract says code and pretrained checkpoints are released. Read the licence text before a commercial program relies on it.

Did OpenVLA-OFT use Firsthand data, and can I request a custom dataset?

No. Firsthand is not affiliated with the OpenVLA-OFT authors and makes no claim that OpenVLA-OFT or any model was trained on Firsthand data. You can request a custom egocentric dataset: write the tasks, fields and format into a spec, confirm it on a Spec Pilot, and billing is per validated hour. Firsthand captures first-person human video, not robot or teleoperation data.