What Is DexVLA?

DexVLA is introduced in arXiv 2502.05855 by Junjie Wen and co-authors, accepted at CoRL 2025 per the abstract page. The abstract says VLA models focus on scaling the vision-language model while action representation stays a bottleneck, and proposes a diffusion-based action expert for cross-embodiment learning. Read: the abstract, full text and GitHub README; the project page rendered no readable text and is not used.

DexVLA 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 paper uses a Qwen2-VL 2B backbone and a diffusion expert built on ScaleDP, a Transformer variant of Diffusion Policy, at one billion parameters. ScaleDP was not designed for cross-embodiment pre-training, so a multi-head output gives each robot configuration its own head. The expert is separable: it is pre-trained alone, then joined to the VLM through a projection module.

Stage 1 pre-trains only the diffusion expert on cross-embodiment data. Stage 2 trains the VLM, projection and expert together on single-embodiment data with the visual encoder frozen. Stage 3 adapts to specific tasks. The authors annotate sub-step instructions, typically every five seconds of a demonstration.

Four configurations: Franka with a gripper, Franka with an Inspire dexterous hand, bimanual UR5e and bimanual AgileX. Tasks include shirt folding, bin picking, table bussing, drink pouring, packing, laundry folding, dryer unloading and sorting. Scores are the authors’ own and are not reproduced here.

The paper credits Diffusion Policy with applying diffusion models to robot learning and builds its expert on ScaleDP. It cites TinyVLA as adding an external action expert, but says such pipelines still rely on the entire model.

DexVLA’s README expects ACT-format h5py files with images, joint states, actions and language; egocentric human video as training data is Not stated in the sources read. Firsthand captures human first-person video, not robot data. 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 DexVLA’s layout is the buyer’s pipeline.

What is DexVLA?

A VLA from arXiv 2502.05855 that pairs a Qwen2-VL 2B backbone with a one-billion-parameter diffusion action expert.

What is the plug-in diffusion expert?

A ScaleDP-based action model with one head per embodiment, pre-trained separately, then joined to the VLM.

How is DexVLA trained?

In three stages: expert pre-training, embodiment alignment and task adaptation, using sub-step annotated data.

Is DexVLA open source?

The GitHub repository releases code, training scripts and ScaleDP-H weights and lists its licence as Other.

Was DexVLA trained on egocentric human video?

Not stated in the sources read. The paper and README describe robot demonstrations.

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

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