Glossary
What Is UniVLA?
Last reviewed: October 2026
Short answer
UniVLA is a framework from a 2025 paper (RSS 2025) for learning cross-embodiment vision-language-action policies from video. It derives task-centric latent actions from video with a latent action model in DINO feature space, using language instructions to separate task-relevant motion, trains a model to predict them, then decodes them to robot actions with a lightweight head.
01
What is UniVLA?
UniVLA is introduced in UniVLA: Learning to Act Anywhere with Task-centric Latent Actions (arXiv 2505.06111). The abstract says most approaches scale action-annotated data and stay tied to one physical specification; UniVLA derives action representations from video across embodiments. Read: the paper, README and univla-7b model card. Other models share the name; this entry covers only this paper.
02
Where does Firsthand fit next to UniVLA?
UniVLA is third-party research. Firsthand is not affiliated with its authors and makes no claim that any named model was trained on, used with or is compatible with Firsthand data. Firsthand captures first-person human video, fast, on demand and to a written spec, not robot data: 3D hand pose at 21 keypoints per hand with a per-joint visibility flag, wrist trajectories and action segments. The only language field is one task instruction per episode.
03
What are task-centric latent actions?
Per the paper, a latent action model encodes the change between two frames about one second apart as discrete tokens, using DINOv2 features and a VQ-VAE objective. Stage 1 is conditioned on a T5 instruction embedding, so its latent actions absorb task-irrelevant change such as other agents and camera motion. Stage 2 freezes that codebook and learns a new one for task-centric latent actions.
04
How does UniVLA learn across embodiments from video?
The paper labels any video frame with latent actions, so robot demonstrations and egocentric human video share one action space without manual alignment. A Prismatic-7B vision-language model predicts the latent action tokens autoregressively.
05
What are UniVLA’s training stages?
There are three. Task-centric latent action learning runs in the two steps above. Next-latent action prediction trains the vision-language model on those latent actions as pseudo-labels. Latent decoding adds embodiment-specific action decoder heads and fine-tunes with LoRA on downstream data.
06
How does UniVLA compare with LAPA and other latent action approaches?
The paper names LAPA and IGOR as earlier unsupervised latent action pretraining and says they encode all visual change from raw pixels, including task-irrelevant dynamics such as camera shake. UniVLA answers with DINOv2 features and a two-stage split, and reproduces LAPA as a baseline; scores are not reproduced here.
07
What do unlabeled videos not capture for robots?
The paper says UniVLA is mainly evaluated on single-arm manipulation, its latent action granularity and codebook size are fixed, and dexterous hands may need finer-grained action modelling. A hand pose input is Not stated in the sources read.
08
How do you get custom egocentric data for cross-embodiment VLA training?
UniVLA is a method, not a data source. For first-person human video of particular tasks, environments or people, write them into a spec and confirm it on a Spec Pilot, a small fast first batch. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where it is thin. Billing is per validated hour; delivery is RLDS, LeRobot, WebDataset or HDF5.
09
UniVLA training stages, inputs and outputs, as stated in the official sources
| Item | What the official sources state |
|---|---|
| Latent action learning, stage 1 | Frame pair and a T5 instruction embedding in; task-irrelevant latent actions out; DINOv2 features, VQ-VAE objective |
| Latent action learning, stage 2 | Stage-1 codebook frozen; a new codebook learns task-centric latent actions |
| Next-latent action prediction | Observation and instruction in; four latent action tokens out; Prismatic-7B backbone |
| Latent decoding | Visual and latent action embeddings in; robot actions out; LoRA fine-tuning, action chunks |
| Training data | Subsets of Open X-Embodiment, GNM and Ego4D (paper) |
| Licence | Apache License 2.0 (GitHub); model card: Not stated in the sources read |
11
UniVLA beside LAPA, as each official source states
| Aspect | UniVLA (sources read) | LAPA (per the site’s entry) |
|---|---|---|
| Action signal | Task-centric latent actions from DINOv2 features, in two stages | Discrete latent actions from frame pairs; the UniVLA paper says raw pixels |
| Language use | Instruction embedding conditions stage 1; policy takes the instruction | A task description per sample |
| Backbone | Prismatic-7B | 7B Large World Model (LWM-Chat-1M) |
| Human video | Ego4D subset | Something-Something V2 |
| Robot step | Embodiment-specific action decoder heads, LoRA | New action head on labeled delta end-effector trajectories |
12
What UniVLA consumes beside what the Firsthand site states
| Aspect | UniVLA (sources read) | What the Firsthand site states |
|---|---|---|
| Visual input | Observation frames; frame pairs about one second apart | Head and wrist camera video |
| Hand data | No hand pose input described | 21 keypoints per hand, visibility flag, wrist trajectories, action segments |
| Language | An instruction per sample; mostly short-horizon | One task instruction per episode |
| Robot actions | Decoder heads trained on downstream low-level actions | Human video, not robot data |
| Compatibility | Not claimed | Not claimed |
13
Frequently asked questions
- What is UniVLA?
- A framework from a 2025 paper that learns cross-embodiment vision-language-action policies from video using task-centric latent actions.
- Does UniVLA need robot action labels?
- Not for latent action learning or pretraining on video. Downstream adaptation trains an action decoder with low-level robot actions.
- What data does UniVLA learn from?
- The paper trains the latent action model on subsets of Open X-Embodiment, GNM and Ego4D human video.
- Is UniVLA open source?
- GitHub shows the Apache License 2.0. The model card licence is Not stated in the sources read. Not legal advice.
- What results does the paper report?
- The abstract reports state-of-the-art results on manipulation and navigation benchmarks and real robots, using under 1/20 of OpenVLA’s pretraining compute. Other numbers are not reproduced here.
- Does Firsthand use UniVLA, and can I request a custom dataset?
- Firsthand is not affiliated with the UniVLA authors and makes no claim of use, training or compatibility. You can request a custom egocentric dataset: write a spec, confirm it on a Spec Pilot, pay per validated hour.
Related terms