What Is TinyVLA?

TinyVLA is introduced in arXiv 2409.12514, with an official project page and GitHub repository. The abstract says current VLA models are slow at inference and need extensive robot pre-training, which hinders deployment. Read: the abstract, project page and README; not the full paper or Hugging Face pages.

TinyVLA is third-party research. Firsthand is not affiliated with its authors and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract names two components: a policy backbone initialized from robust, high-speed multimodal models, and a diffusion policy decoder integrated during fine-tuning. The README says the backbone puts the Pythia language model into the LLaVA framework, with Llava-Pythia weights of about 400M, 700M and 1.3B parameters for TinyVLA-S, -B and -H.

The abstract says TinyVLA eliminates the pre-training stage, so the policy is fine-tuned on task data. The README expects ACT-format h5py files, with a script to convert RLDS. Its example holds actions, a raw language entry, left, right and wrist images, joint positions, qpos and qvel. Demonstration counts are Not stated in the sources read.

The abstract names slow inference and large robot pre-training as barriers to deployment, and claims faster inference and better data efficiency than OpenVLA with comparable or superior performance. Speed figures are Not stated in the sources read, and scores are not reproduced.

All four target robot control but differ in backbone, action output and training data. The table gives only what each official source or site entry states; Firsthand ranks none.

The sources describe robot demonstrations with joint states. Egocentric human video as training data is not described in the sources read, and dataset sizes are Not stated. Coverage of tasks, environments and people belongs in a written data spec.

Firsthand captures human first-person video, not robot data, and states no robot joint state or actions. Write tasks, environments and signals into a spec and confirm them on a Spec Pilot. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where thin. Formats: RLDS, LeRobot v2.1, WebDataset, HDF5; mapping to TinyVLA’s layout is the buyer’s pipeline.

What is TinyVLA?

A family of compact VLA models for robot manipulation with a fast multimodal backbone and a diffusion policy decoder.

Does TinyVLA need robot pretraining?

The abstract says it eliminates the pre-training stage. Dataset sizes are Not stated in the sources read.

How is TinyVLA different from OpenVLA?

The abstract claims faster inference and better data efficiency, from the authors’ own tests. Scores are not reproduced.

Is TinyVLA open source?

The repository lists the MIT License. Model weight licences are Not stated in the sources read.

Was TinyVLA trained on egocentric human video?

Not described in the sources read. The README describes robot demonstrations with joint states.

Does Firsthand use TinyVLA, and can I request a custom dataset?

Firsthand is not affiliated with the TinyVLA authors and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.