Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Official repo for paper DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.
| Date | Stars |
|---|---|
| 2026-07-31 | 393 |
| 2026-08-01 | 393 |
| 2026-08-02 | 393 |
| 2026-08-06 | 393 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<h3 align="center">
🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉<br>
<font color="red"><b>Check out our latest progress</b></font> of new offline RL algorithm for Android <a href="https://digirl-agent.github.io/DigiQ-agent.github.io/"><b>DigiQ</b></a> and autonomous skill discovery for web agents <a href="https://yanqval.github.io/PAE/"><b>Proposer-Agent-Evaluator</b></a>. <br>
🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉🎉
</h3>
<p align="center">
<img src="./assets/digirl-logo-text.png" alt="logo" width="50%">
</p>
<h3 align="center">
DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning
<br>
<b>Oral @ <a href="https://icml-fm-wild.github.io/">FM Wild</a>, ICML</b>
<br>
<b>Neurips 2024</b>
</h3>
<p align="center">
| <a href="https://digirl-agent.github.io/"><b>Website | Demo | Results</b></a> | <a href="https://arxiv.org/abs/2406.11896"><b>Paper</b></a> | <a href="https://drive.google.com/drive/folders/14Iu6lAHePQ2qG0ghYkVG1RG6RUu7e2Hz?usp=sharing"><b>Checkpoints | Data</b></a> |
</p>
---
Research Code for preprint "DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning".
[Hao Bai*](https://jackgethome.com), [Yifei Zhou*](https://<username>02.github.io/), [Mert Cemri](https://scholar.google.com/citations?user=sMEFwf8AAAAJ&hl=en), [Jiayi Pan](https://www.jiayipan.me/), [Alane Suhr](https://www.alanesuhr.com/), [Sergey Levine](https://people.eecs.berkeley.edu/~svlevine/), [Aviral Kumar](https://aviralkumar2907.github.io/)<br>
UC Berkeley, UIUC, Google DeepMind
<br>
*Equal contribution, alphabetic order; work done at UC Berkeley
<p align="center">
<img src="./assets/digirl-diagram.png" alt="digirl-diagram" width="70%">
</p>
## 🍩 Features
### Environment Features
- Auto-adaptive error handling support.
- Multi-machine [emulation parallel](multimachine/README.md) support.
- Checkpoint resuming support.
- Trajectory video recording support.
### Approach Features
- Two training algorithms proposed in the paper
- DigiRL (automatic curriculum + doubly robust estimator filtering).
- Filtered Behavior Cloning (reward-based filtering).
- Three training modes:
- Offline-only training: baseline apporach - use the AutoUI checkpoint to collect data (we have this data ready for you), then train with these pre-collected sub-optimal trajectories. This mode only allows evaluation using the checkpoint.
- Online-only training: traditional RL approach - the AutoUI checkpoint simultaneously interacts with the environment learns online. This mode allows interactive training.
- Offline-to-online training: the most powerful approach as evaluated in paper - the AutoUI checkpoint first learns the pre-collected data, then simultanesouly interacts with the environment and do online learning starting from this checkpoint. This mode allows interactive training
- Two agents:
- [AutoUI](https://arxiv.org/abs/2309.11436): we support both training (2 algorithms x 3 paradigms) and evaluation.
- [CogAgent](https://arxiv.org/abs/2312.08914): current only support evaluation, no training pipeline is supported.
- Two [Android-in-the-Wild](https://arxiv.org/abs/2307.10088) task sets:
- AitW General: general browsing, opening apps.
- AitW Web Shopping: shopping on popular shopping websites.
- It'll also be interesting to explore the [other AitW subsets](https://github.com/google-research/google-research/tree/master/android_in_the_wild) or other task sets if you have good candidates, please propose one in the issue.
- DDP Multi-GPU training:
- We support `accelerate` for multi-GPU training. You can turn off this feature if you only have 1 GPU. It only takes **12GB** of GPU memory for AutoUI running the DigiRL algorithm, but we provide this feature in case you want to play with something larger.
## 🚀 Quick Start
### Dependencies
First, create a [conda](https://conda.io/projects/conda/en/latest/user-guide/install/index.html) environment and install all pip package requireExcerpt of 12,586 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:b75d5cd04dc96440, llm:Repository description: 'Official repo for paper DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.'
matched fp:b75d5cd04dc96440, llm:Repository description: 'Official repo for paper DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.'
matched fp:b75d5cd04dc96440, llm:Repository description: 'Official repo for paper DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.'