Robo-Dopamine-GRM-2.0-4B-Preview/README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen3-VL-4B-Instruct
pipeline_tag: image-text-to-text
---
<h1 align="center">Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation</h1>

<h3 align="center">Joy is dopamine’s handiwork—whether in humans or in robotics.</h3>


<p align="center">
  <a href="https://arxiv.org/abs/2512.23703"><img src="https://img.shields.io/badge/arXiv-2512.23703-b31b1b.svg" alt="arXiv"></a>
  &nbsp;
  <a href="https://robo-dopamine.github.io/"><img src="https://img.shields.io/badge/%F0%9F%8F%A0%20Project-Homepage-blue" alt="Project Homepage"></a>
  &nbsp;
  <a href="https://github.com/FlagOpen/Robo-Dopamine"><img src="https://img.shields.io/badge/🔍%20Code-Github-orange" alt="Github"></a>
  &nbsp;
  <a href="https://huggingface.co/collections/tanhuajie2001/robo-dopamine"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Weights-Huggingface-yellow" alt="Weights"></a>
  &nbsp;
  <a href="#"><img src="https://img.shields.io/badge/🤗%20Dataset-Stay%20tuned-green.svg" alt="Dataset"></a>

</p>


<div style="text-align: center; background-color: white;">
    <img src="https://github.com/FlagOpen/Robo-Dopamine/raw/main/assets/teasor.png" width=100% >
</div>


## 🗞️ News
- **`2026-04-05`**: 🤗 We released [Robo-Dopamine-GRM-2.0-4B-Preview](https://huggingface.co/tanhuajie2001/Robo-Dopamine-GRM-2.0-4B-Preview) model.
- **`2026-03-05`**: 🤗 We released [Robo-Dopamine-GRM-2.0-8B-Preview](https://huggingface.co/tanhuajie2001/Robo-Dopamine-GRM-2.0-8B-Preview) model. More General, More Powerful!!!
- **`2026-03-02`**: 🤗 We released [Robo-Dopamine-GRM-8B](https://huggingface.co/tanhuajie2001/Robo-Dopamine-GRM-8B) model
- **`2026-02-22`**: 🔥🔥🔥 **Robo-Dopamine** gets accepted to CVPR 2026! See you in Denver, Colorado, USA!
- **`2026-02-10`**: ⚡  We released data generation pipeline and finetune codes. ***Try to finetune with your own data***.
- **`2026-01-26`**: 🔍 We released [Robo-Dopamine-Bench](https://huggingface.co/datasets/tanhuajie2001/Robo-Dopamine-Bench) benchmark and evaluation codes.
- **`2026-01-08`**: 🤗 We released [Robo-Dopamine-GRM-3B](https://huggingface.co/tanhuajie2001/Robo-Dopamine-GRM-3B) model and inference codes.
- **`2025-12-30`**: ✨ ***Codes, Dataset and Weights are coming soon! Stay tuned for updates***.
- **`2025-12-30`**: 🔥 We released our [Project Page](https://robo-dopamine.github.io/) of **Robo-Dopamine**.


## 🤖 Overview

**Robo-Dopamine** is composed of two core components: ***(a) Dopamine-Reward Modeling Method --*** At the heart of our reward modeling is to build the General Reward Model (GRM), a vision-language model that is prompted with a task description and conditioned on multi-view images of initial, goal, "BEFORE," and "AFTER" states to predict a relative progress or regress hop. To ensure a stable and accurate signal, we employ *Multi-Perspective Progress Fusion*, which combines incremental, forward-anchored, and backward-anchored predictions into a final fused reward. And ***(b) Dopamine-RL Training Framework --*** The Dopamine-RL framework first adapts the pre-trained GRM to a novel task using a single demonstration, i.e., *One-Shot GRM Adaptation*. Subsequently, it uses a theoretically-sound *Policy-Invariant Reward Shaping* method to convert the GRM's dense output into a reward signal that accelerates learning without altering the optimal policy. 
This approach is universally compatible with a wide range of RL algorithms.

<div align="center"> 
    <img src="https://github.com/FlagOpen/Robo-Dopamine/raw/main/assets/method.png" alt="Logo" style="width=100%;vertical-align:middle">
</div>

## 🛠️ Setup

```bash
# clone repo.
git clone https://github.com/FlagOpen/Robo-Dopamine.git
cd Robo-Dopamine

# build conda env.
conda create -n robo-dopamine python=3.10
conda activate robo-dopamine
pip install -r requirements.txt
```

## 💡 Simple Inference

```python
import os
from examples.inference import GRMInference

model = GRMInference("tanhuajie2001/Robo-Dopamine-GRM-2.0-4B-Preview")

TASK_INSTRUCTION = "organize the table"
BASE_DEMO_PATH = "./examples/demo_table"
OUTPUT_ROOT = "./results"

## Note: If no target/goal image is provided, 
## please replace `GOAL_IMAGE_PATH` with the blank image "./examples/demo_table/blank_goal.png".
GOAL_IMAGE_PATH = "./examples/demo_table/goal_image.png" # "./examples/demo_table/blank_goal.png"

# select prediction model: Forward-Mode, Incremental-Mode or Backward-Mode
PREDICTION_MODE = "forward" # "incremental" or "backward"

# multi-view usage:
output_dir = model.run_pipeline(
    cam_high_path  = os.path.join(BASE_DEMO_PATH, "cam_high.mp4"),
    cam_left_path  = os.path.join(BASE_DEMO_PATH, "cam_left_wrist.mp4"),
    cam_right_path = os.path.join(BASE_DEMO_PATH, "cam_right_wrist.mp4"),
    out_root       = OUTPUT_ROOT,
    task           = TASK_INSTRUCTION,
    frame_interval = 10, # modify frame_interval as desired, but it shouldn't be set too small if using 'incremental'.
    batch_size     = 1, # please increase batch_size > 1, if you have enough GPU memory.
    goal_image     = GOAL_IMAGE_PATH,
    eval_mode      = PREDICTION_MODE,
    visualize      = True
)
print(f"Episode ({BASE_DEMO_PATH}) processed with multi-view {PREDICTION_MODE}-mode. Output at: {output_dir}")

# single-view usage:
output_dir = model.run_pipeline(
    cam_high_path  = os.path.join(BASE_DEMO_PATH, "cam_high.mp4"),
    cam_left_path  = os.path.join(BASE_DEMO_PATH, "cam_high.mp4"), # repeat cam_high
    cam_right_path = os.path.join(BASE_DEMO_PATH, "cam_high.mp4"), # repeat cam_high
    out_root       = OUTPUT_ROOT,
    task           = TASK_INSTRUCTION,
    frame_interval = 10, # modify frame_interval as desired, but it shouldn't be set too small if using 'incremental'.
    batch_size     = 1, # please increase batch_size > 1, if you have enough GPU memory.
    goal_image     = GOAL_IMAGE_PATH,
    eval_mode      = PREDICTION_MODE,
    visualize      = True
)
print(f"Episode ({BASE_DEMO_PATH}) processed with single-view {PREDICTION_MODE}-mode. Output at: {output_dir}")

```

## 📑 Citation

If you find our work helpful, feel free to cite it:
```
@article{tan2025robo,
  title={Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation},
  author={Tan, Huajie and Chen, Sixiang and Xu, Yijie and Wang, Zixiao and Ji, Yuheng and Chi, Cheng and Lyu, Yaoxu and Zhao, Zhongxia and Chen, Xiansheng and Co, Peterson and others},
  journal={arXiv preprint arXiv:2512.23703},
  year={2025}
}
```