What Is VLA-Adapter?

VLA-Adapter is introduced in arXiv 2509.09372 by Yihao Wang and co-authors, with an official project page, GitHub repository and Hugging Face organization. The abstract says VLA models typically bridge perception and action by pre-training a large vision-language model on robotic data, which raises performance but also training cost. Read: the abstract, project page, most of the README and the Hugging Face page; not the full paper text.

VLA-Adapter is third-party research. Firsthand is not affiliated with its authors and claims no use, training or compatibility with it. Firsthand captures real 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 training cost: large-scale VLM pre-training on robotic data helps, but is expensive. The project page asks whether shrinking the backbone is enough, compares OpenVLA-OFT with 7B and 0.5B backbones, and concludes that an effective way to bridge vision-language to action is needed to cut backbone size while keeping performance. Its result tables are not reproduced here.

The project page says the VLM follows the Prismatic-VLMs architecture. At each layer, Raw features (the VLM’s vision and language representations) and ActionQuery features (additional learnable tokens) are integrated by Bridge Attention with the matching action latent, and the Policy has the same layer count as the VLM. The degree of Raw-feature injection is learnable. Total model size is Not stated in the sources read.

The abstract says no robotic data pre-training, and a powerful VLA trained in 8 hours on a single consumer-grade GPU. The README fine-tunes a Prismatic-VLMs Qwen2.5-0.5B checkpoint with LoRA, lists configurations for GPUs from 10GB to 80GB or more of VRAM, and says the paper used four H100 GPUs. The project page says the backbone can also be frozen, training only the ActionQuery latent and Policy.

The abstract says experiments on simulated and real-world robotic benchmarks show state-of-the-art-level performance and fast inference. The project page reports the LIBERO benchmark (Spatial, Object, Goal and Long suites), CALVIN ABC→D, and a throughput comparison with OpenVLA and OpenVLA-OFT. Scores and speeds are not reproduced here; real-world task details are Not stated in the sources read.

The sources read describe robot benchmarks: the README downloads LIBERO and CALVIN data in RLDS, and its example configuration uses two images per input plus proprioception. Egocentric human video as training data is not described. Firsthand states no robot actions or joint state, so it is not a drop-in swap for that data; any use with a given policy is the buyer’s pipeline.

Firsthand captures real human first-person video, with no synthetic or simulated 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.

What is VLA-Adapter?

A paradigm for tiny-scale VLA models with a lightweight Policy module and Bridge Attention, introduced in arXiv 2509.09372.

Does VLA-Adapter need robotic data pre-training?

The abstract says it reaches high performance with a 0.5B-parameter backbone and no robotic data pre-training.

How long does VLA-Adapter take to train?

The abstract says 8 hours on a single consumer-grade GPU. Other training times are Not stated in the sources read beyond the README’s per-configuration notes.

Is VLA-Adapter open source?

The repository lists the MIT License, and checkpoints are on the Hugging Face organization. Weight licences are Not stated in the sources read.

Was VLA-Adapter trained on egocentric human video?

Not described in the sources read. The README describes LIBERO and CALVIN robot benchmark data.

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

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