---
license: other
license_name: rlwrld-model-license-v1.0
license_link: LICENSE.md
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - vlm
  - vision-language
  - rldx
  - qwen3-vl
base_model: Qwen/Qwen3-VL-8B-Instruct
---

# RLDX-1-VLM

[Paper](https://arxiv.org/abs/2605.03269)  ·  [Project page](https://rlwrld.ai/rldx-1)  ·  [Code](https://github.com/RLWRLD/RLDX-1)  ·  [Models](https://huggingface.co/collections/RLWRLD/rldx-1)

`RLDX-1-VLM` is the **vision-language backbone** used by the RLDX-1 robot
policy family. It is a Qwen3-VL-8B-Instruct checkpoint distributed
separately from the action policy so that researchers can inspect, finetune,
or replace the perceptual stack independently.

> **Note.** This checkpoint exposes a standard Qwen3-VL VLM interface only —
> it does **not** ship the Multi-Stream Action Transformer head, the
> cognition tokens, the memory / motion / physics modules, or the RLDX
> inference server. For action prediction, use one of the
> [`RLDX-1-PT`](https://huggingface.co/RLWRLD/RLDX-1-PT),
> [`RLDX-1-FT-*`](https://huggingface.co/RLWRLD), or
> [`RLDX-1-MT-*`](https://huggingface.co/RLWRLD) checkpoints.

## Intended use

- As the `--backbone-path` for finetuning a fresh RLDX-1 policy ([recipe](https://github.com/RLWRLD/RLDX-1#finetuning)).
- For VLM-only ablations, dense-captioning experiments, or perceptual
  probing within the RLDX research stack.

## Quick start

```python
from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "RLWRLD/RLDX-1-VLM",
    torch_dtype="bfloat16",
    device_map="cuda:0",
)
processor = AutoProcessor.from_pretrained("RLWRLD/RLDX-1-VLM")
```

## Model details

- **Type:** Vision-language model (multimodal text + image / video).
- **Backbone:** [`Qwen/Qwen3-VL-8B-Instruct`](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct).
- **Params:** 8B.
- **Role in RLDX-1:** perceptual encoder for the MSAT action policy. Cognition
  tokens are injected into this backbone and routed through Qwen3-VL hidden
  states to produce a compact perceptual summary consumed by the action
  model.

For a full architectural walkthrough including how cognition tokens are
wired into this backbone, see
[`docs/architecture.md`](https://github.com/RLWRLD/RLDX-1/blob/main/docs/architecture.md).

## Limitations

`RLDX-1-VLM` is a research backbone snapshot. It is not safety-tuned beyond
its Qwen3-VL upstream, and it is not intended as a general-purpose
chat assistant. For action prediction it must be paired with the RLDX-1
policy head; the standalone VLM does not produce robot commands.

## Citation

```bibtex
@article{rldx2026,
  title={RLDX-1 Technical Report},
  author={Kim, Dongyoung and Jang, Huiwon and Koo, Myungkyu and Jang, Suhyeok and Kim, Taeyoung and others},
  year={2026},
  note={RLWRLD},
  eprint={2605.03269},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2605.03269}
}
```

## License

Released under the **RLWRLD Model License v1.0** — a non-commercial license
with attribution and share-alike requirements. See [`LICENSE.md`](LICENSE.md) for
the full text. By using this model you agree to those terms, including the
use restrictions in §3.5.