dfexpense

KO | EN

gitlite — search

#javascript #hacktoberfest #go #golang #nodejs #java #windows #awesome #awesome-list #vue #pytorch #openai #ios #mcp

(★ 10)

Code for "Diffusion Filtered Exploration via Ensembles (DF-ExpEnse)" (ICML 2026)

dfexpense 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h2><center>Diffusion Filtered Exploration via Ensembles (DF-ExpEnse)</h2>

[Calvin Luo](calvinyluo.com) (🌲🧸), [Chen Sun](chensun.me) (🧸), [Shuran Song](shurans.github.io) (🌲)

ICML 2026

[[Website](https://df-expense.github.io)] [[Paper](http://arxiv.org/abs/2606.19656)]
</div>

## Overview
DF-ExpEnse is an exploration technique for collecting high-quality experience during online interaction, thus improving sample-efficiency when performing RL finetuning of pretrained generative control policies.

DF-ExpEnse is general, as it only adjusts online experience collection, and can therefore be integrated on top of a variety of update techniques.  As such, the codebase and setup will largely follow that of the chosen underlying RL finetuning strategy - in this instantiation, [DSRL](https://github.com/ajwagen/dsrl).

## Installation
1. Clone Repository
```
git clone git@github.com:real-stanford/dfexpense.git
cd dfexpense
```
2. Create Conda Environment
```
conda create -n dfexpense python=3.9 -y
conda activate dfexpense
```
3. Install DPPO Package
```
cd dppo
pip install -e .
pip install -e .[robomimic]
pip install -e .[gym]
cd ..
```
4. Install Stable Baselines3 Package with DF-ExpEnse Modifications
```
cd stable-baselines3
pip install -e .
cd ..
```
5. Download Pretrained BC Checkpoints and Normalization Statistics
```
python scripts/pretrained_bc_setup.py
```

## Running DF-ExpEnse
To run DF-ExpEnse on Robomimic, call
```
python train.py --config-path=cfg/robomimic --config-name=square.yaml
```
To run DF-ExpEnse on Gym, call
```
python train.py --config-path=cfg/gym --config-name=walker.yaml
```

## Acknowledgements
This implementation of DF-ExpEnse is integrated on top of [DSRL](https://github.com/ajwagen/dsrl), which is itself built on top of [Stable Baselines3](https://github.com/DLR-RM/stable-baselines3). For most diffusion policy implementations, we continue to utilize the implementation given in the [DPPO](https://github.com/irom-princeton/dppo) codebase, as done in DSRL.  For the Tool Hang implementation, we use the [DICE-RL](https://github.com/real-stanford/dice-rl/) flow matching [implementation](https://github.com/real-stanford/dice-rl/tree/main/model/flow_matching) and [checkpoint](https://huggingface.co/wintermelontree/robomimic-pretrain-checkpoints/tree/main/tool_hang_low_dim/checkpoint).

## Citation
```
@inproceedings{luo2026dfexpense,
      title={{DF}-ExpEnse: Diffusion Filtered Exploration for Sample Efficient Finetuning},
      author={Calvin Luo and Chen Sun and Shuran Song},
      booktitle={Forty-third International Conference on Machine Learning},
      year={2026}
    }
```

dfexpense

dfexpense 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

dfexpense

dfexpense 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README