README.md

---
title: DeepThinkVLA
emoji: 🤖
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.5.0
python_version: "3.10"
app_file: app.py
pinned: false
license: mit
short_description: Reason then act — CoT + 7-DoF robot action chunks
startup_duration_timeout: 1h
---

# DeepThinkVLA

Demo of [**DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action
Models**](https://huggingface.co/papers/2511.15669) (arXiv 2511.15669).

DeepThinkVLA is a ~3B PaliGemma-based Vision-Language-Action model trained with a
two-stage SFT + RL recipe so that its chain-of-thought is *causally useful* for the
action it emits. Given an agent-view image, a wrist-camera image and a task
instruction, it first writes out a reasoning trace and then decodes a **10-step,
7-DoF action chunk** in a single non-autoregressive, bidirectionally-attended pass.

* Checkpoint: [`yinchenghust/deepthinkvla_libero_cot_rl`](https://huggingface.co/yinchenghust/deepthinkvla_libero_cot_rl) (LIBERO CoT SFT + RL)
* Upstream code: [OpenBMB/DeepThinkVLA](https://github.com/OpenBMB/DeepThinkVLA) (MIT)

## Fidelity notes

The inference path mirrors the authors' `src/experiments/deepthinkvla_utils.py::get_vla_action`
exactly: same `<image><image>` + `THINK_PREFIX` + `Task: …;` prompt, same greedy
generation with the action-token ban and the `</think><action>` stopping criterion,
same `prompt_cot_predict_action` action head, same QUANTILE de-normalization from
`norm_stats.json`, same gripper binarization before execution. The custom modeling
code (`sft/`, `dt_datasets/`) is vendored verbatim from the upstream repo. The only
deviation is that the 256→224 resize uses `PIL.Image.resize(..., BILINEAR)` directly
instead of `torchvision.transforms.Resize`, which dispatches to the identical PIL call.

The Space performs single-step prediction on static observations — it does not run the
LIBERO simulator, so there is no closed-loop rollout.

## Example assets

Example observation pairs are frames from the authors' own
[`yinchenghust/libero_cot`](https://huggingface.co/datasets/yinchenghust/libero_cot)
dataset (Apache-2.0). They are stored in the orientation the policy is trained on
(the standard OpenVLA LIBERO 180° rotation), which is why they look mirrored.

`LICENSE_DeepThinkVLA` is the MIT license of the vendored upstream code.