Glossary
What Is OneTwoVLA?
Last reviewed: October 2026
Short answer
OneTwoVLA is a published vision-language-action model, introduced in arXiv 2505.11917, that performs both acting (System One) and reasoning (System Two) in a single model. It adaptively switches between reasoning at critical moments and acting on the most recent reasoning, and is co-trained with synthetic vision-language data. It is third-party research, not Firsthand’s.
01
What is OneTwoVLA?
OneTwoVLA is introduced in arXiv 2505.11917 by Fanqi Lin, Ruiqian Nai, Yingdong Hu, Jiacheng You, Junming Zhao and Yang Gao, with an official project page and GitHub repository. The abstract says general-purpose robots need synergistic reasoning and acting. Read: the abstract, the project page and the README; not the full paper text. The Hugging Face datasets the README mentions were not opened.
02
Where does Firsthand fit next to OneTwoVLA?
OneTwoVLA 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.
03
What problem does OneTwoVLA address, per the sources?
The abstract says recent dual-system approaches, which separate high-level reasoning from low-level acting, often suffer from limited mutual understanding of capabilities between the two systems and from latency. OneTwoVLA is offered as a single model that does both, so reasoning and acting live in one set of weights rather than two cooperating systems.
04
How does OneTwoVLA decide when to reason and when to act, per the sources?
The abstract says the model adaptively switches between two modes: explicitly reasoning at critical moments during task execution, and generating actions based on the most recent reasoning at other times. The project page adds that it can adjust actions immediately on human intervention and proactively seek clarification when faced with ambiguity. The internal switching mechanism and backbone are Not stated in the sources read.
05
How is OneTwoVLA trained, per the sources?
The abstract says the authors designed a scalable pipeline for synthesizing embodied reasoning-centric vision-language data, used for co-training with robot data. The project page says the pipeline is automatic, without artificial intervention, and covers visual grounding and long-horizon planning instructions. The README lists training scripts for cocktail and visual grounding tasks and says its code builds on openpi. Training time and compute are Not stated in the sources read.
06
How was OneTwoVLA evaluated, per the sources?
The abstract reports experiments on four capabilities: long-horizon task planning, error detection and recovery, natural human-robot interaction, and generalizable visual grounding. The project page names tasks such as hotpot cooking, tomato-egg scramble and cocktail mixing, and unseen objects in grounding. Scores and success rates are not reproduced here and are Not stated in the sources read beyond those descriptions.
07
How does a VLA relate to real human egocentric video?
The sources read describe robot data and synthetic vision-language images. The README says the robot datasets use UMI data converted to the LeRobot format, and it augments synthetic images to resemble real robot observations. 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; any use with a given policy is the buyer’s pipeline.
08
What real-world egocentric demonstration data can you get custom-collected on demand?
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.
09
OneTwoVLA architecture, reasoning modes and training setup, as stated in the official sources
| Item | What the official sources state |
|---|---|
| Publication | arXiv 2505.11917; a second version is dated 1 March 2026 on arXiv |
| Model type | Single unified VLA performing acting (System One) and reasoning (System Two) |
| Reasoning mode | Explicit reasoning at critical moments during task execution (abstract) |
| Acting mode | Actions generated from the most recent reasoning at other times (abstract) |
| Contrast | Dual-system approaches separate high-level reasoning from low-level acting (abstract) |
| Backbone and size | Not stated in the sources read; the code builds on openpi (README) |
| Co-training | Robot data plus synthetic embodied reasoning-centric vision-language data |
| Image augmentation | Random fisheye distortion and a composited robot gripper with adaptive brightness (README) |
| Robot data format | UMI data, with conversion code to LeRobot format (README) |
| Code | MIT License on the repository; weight licences Not stated in the sources read |
| Firsthand relationship | None |
11
Capabilities described on the OneTwoVLA project page
| Capability | What the project page states |
|---|---|
| Long-horizon task planning | Reasons to formulate, track and adjust plans from execution feedback; examples include hotpot cooking and cocktail mixing |
| Error detection and recovery | Detects errors in real time, reasons about recovery and performs corrective actions |
| Human-robot interaction | Adjusts on human intervention and seeks clarification when instructions are ambiguous |
| Visual grounding | Understands spatial relationships, object attributes and semantic features, including unseen objects |
| Synthetic data types | Visual grounding tasks and long-horizon planning tasks, each with written reasoning |
12
What OneTwoVLA’s sources describe beside what the Firsthand site states
| Aspect | OneTwoVLA (sources read) | What the Firsthand site states |
|---|---|---|
| Robot data | Robot datasets for cocktail and visual grounding tasks (README) | Not stated; human video, not robot data |
| Vision-language data | Synthetic images with annotated reasoning (README) | Real human capture only; no synthetic or simulated data |
| Reasoning text | Written reasoning per instruction on the project page | One task instruction per episode, plus action segments |
| Images | Synthetic images augmented to resemble real robot observations | Head-mounted egocentric video plus two wrist cameras |
| Hands | Not described in the sources read | 21 keypoints per hand with a per-joint visibility flag, plus wrist trajectories |
| File layout | UMI converted to LeRobot format (README) | RLDS, LeRobot v2.1, WebDataset, HDF5 |
| Compatibility | Not claimed | Not claimed |
13
Frequently asked questions
- What is OneTwoVLA?
- A single unified vision-language-action model that performs both acting and reasoning and adaptively switches between them, introduced in arXiv 2505.11917.
- When does OneTwoVLA reason instead of act?
- Per the abstract, it reasons explicitly at critical moments during task execution and generates actions from the most recent reasoning at other times.
- What data was OneTwoVLA co-trained with?
- Robot data plus synthetic embodied reasoning-centric vision-language data, per the abstract. Dataset sizes are Not stated in the sources read.
- Is OneTwoVLA open source?
- The repository lists the MIT License and provides training scripts and datasets hosted on Hugging Face. Weight licences are Not stated in the sources read.
- Was OneTwoVLA trained on egocentric human video?
- Not described in the sources read. The README describes UMI robot data and synthetic vision-language images.
- Does Firsthand use OneTwoVLA, and can I request a custom dataset?
- Firsthand is not affiliated with the OneTwoVLA authors and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.
Related terms