What Is Diffusion-VLA?

Diffusion-VLA, also written DiffusionVLA and abbreviated DiVLA, is introduced in Diffusion-VLA: Generalizable and Interpretable Robot Foundation Model via Self-Generated Reasoning (arXiv 2412.03293; v1 4 Dec 2024, v3 4 Jun 2025) by Junjie Wen, Minjie Zhu, Yichen Zhu and eight co-authors. The arXiv page lists ICML 2025, and the project page lists Midea Group, East China Normal University and Shanghai University. Read: the arXiv abstract page, the official project page and the GitHub README of the repository the project page points to for DiVLA training code. The full paper text was not read.

Third-party research; Firsthand is not affiliated 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.

Per the abstract, Diffusion-VLA combines an autoregression model with a diffusion model for learning a visuomotor policy. Its central piece is a next-token prediction objective, which lets the model reason over the user’s query in the context of current observations. A diffusion model is then attached to generate the action outputs. The project page and the repository BibTeX carry an earlier title, "Scaling Robot Foundation Models via Unified Diffusion and Autoregression", which is where the phrase comes from; the current arXiv title names self-generated reasoning instead. Both point to arXiv 2412.03293.

The abstract introduces a reasoning injection module that integrates reasoning phrases directly into the policy learning process. It says the reasoning module makes the model interpretable, because observers can see the model’s thought process and identify potential causes of policy failures. On the project page, example instructions such as "Sort all the items on the right panel" are shown with the model’s generated reasoning marked in blue. The repository README lists a substep_reasonings field beside the raw language in its example episode layout. How the reasoning phrases are authored is Not stated in the sources read.

The abstract says a diffusion model is attached after the reasoning step to generate robust action outputs. The site’s diffusion policy entry describes the general idea: start from random noise and, conditioned on observations, remove noise over steps to produce future actions. The README adds that DiVLA uses a UNet-based diffusion policy as its policy head and has no three-stage training recipe, unlike DexVLA. Denoising steps, action chunk length and the vision-language backbone for DiVLA are Not stated in the sources read.

The abstract reports tests on multiple real robots, including a factory sorting task where the model categorizes objects, including some not seen during training, and a zero-shot bin-picking task with 63.7% accuracy on 102 previously unseen objects. It also claims robustness to distractors and new backgrounds, easy adaptation to new embodiments, and the ability to follow novel instructions while retaining conversational ability. On efficiency, it says DiVLA-2B runs at 82 Hz on a single A6000 GPU and can train from scratch on fewer than 50 demonstrations for a complex task, and that scaling from 2B to 72B parameters improved generalization. The project page names Diffusion Policy, Octo, TinyVLA and OpenVLA as the factory sorting comparisons and says DiVLA outperforms the runner-up, OpenVLA, by 20.9%. These are the authors’ claims, not Firsthand results.

The site’s entries describe a diffusion policy as a policy class that generates actions by iterative denoising, HybridVLA as embedding diffusion denoising inside next-token prediction in one LLM and fusing both predictions, and DexVLA as a vision-language model with a plug-in diffusion action expert, scaled to one billion parameters and trained in three stages. Diffusion-VLA’s stated move is to reason autoregressively first, attach a diffusion head for the action, and inject reasoning phrases into policy learning. The README says DexVLA is built on DiVLA. The table ranks none of them.

The README says its data format is the same as ACT, converted to h5py, and points to a script that converts from RLDS to h5py. Its example layout holds an action array, the raw language, substep reasonings, multi-view images named left, right and wrist, and joint positions. Firsthand’s site states human video with depth derived from a calibrated stereo rig, hand pose, wrist trajectories and action segments. Robot actions and reasoning phrases are Not stated as deliverables. Whether human data can be mapped into that layout is the buyer’s to check.

The full paper was not read, so the DiVLA backbone, how reasoning phrases are produced, denoising settings, training data and the baseline evaluation protocols are Not stated here. The README belongs to the DexVLA repository, so its DiVLA-specific content is limited to the points above. GitHub lists that repository’s licence as Other, so read its terms before use. Quote the paper for exact protocols.

The sources read describe robot demonstrations, not human video. Firsthand records first-person human video to a written spec: depth from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, action segments and one task instruction per episode. That is the site’s only language data; reasoning phrases and robot actions are Not stated. Confirm tasks and signals on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS; matching DiVLA’s expected fields is the buyer’s to check.

What is Diffusion-VLA?

A vision-language-action model (arXiv 2412.03293) that combines autoregressive next-token reasoning with a diffusion model that generates robot actions, plus a reasoning injection module.

What is self-generated reasoning in DiVLA?

Per the abstract, a reasoning injection module integrates reasoning phrases into policy learning, which makes the model interpretable and helps observers identify potential causes of policy failures.

How does the diffusion action head work in Diffusion-VLA?

A diffusion model is attached after the reasoning step to generate actions. The README says DiVLA uses a UNet-based diffusion policy as its head. Denoising settings are Not stated in the sources read.

How is Diffusion-VLA different from HybridVLA, diffusion policy and DexVLA?

It reasons autoregressively and then attaches a diffusion head. The site entries describe HybridVLA as embedding denoising inside next-token prediction, diffusion policy as a policy class, and DexVLA as a plug-in diffusion expert. The README says DexVLA is built on DiVLA.

Is Diffusion-VLA open source?

The project page says the DexVLA code can also train DiVLA, and the README names a DiVLA training script. GitHub lists the repository licence as Other, so read its terms.

Does Firsthand use Diffusion-VLA, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. The site states one task instruction per episode and no reasoning phrases or robot actions. You can request a custom dataset, confirmed first on a Spec Pilot.