Awesome AI Agents › Research Platforms & Simulators

DigiRL-agent/digirl

⭐ 396 Python added to this list on 2025-04-19 repository created 2024-06-20

DigiRL is a research project focused on training device-control agents using autonomous reinforcement learning in real-world environments. The project provides code and resources for the paper "DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning," presented at ICML and NeurIPS 2024. The repository supports training and evaluation of agents that can control Android devices through reinforcement learning techniques. It features two main training algorithms: DigiRL, which combines automatic curriculum learning with doubly robust estimator filtering, and Filtered Behavior Cloning, which uses reward-based filtering. The project supports three training modes: offline-only training using pre-collected sub-optimal trajectories, online-only training where the agent learns interactively, and offline-to-online training that combines both approaches for enhanced performance. The repository includes two agents: AutoUI, which supports both training and evaluation, and CogAgent, which currently supports evaluation only. It also provides two Android-in-the-Wild task sets for general browsing and web shopping tasks. The environment supports multi-machine emulation, checkpoint resuming, trajectory video recording, and multi-GPU training with efficient memory usage. Users can set up the environment with detailed instructions, including dependencies, Android environment setup, and downloading pre-trained model checkpoints and datasets. The project offers pre-collected trajectory data and final offline-to-online checkpoints for reproducing results. Configuration files allow customization of training parameters, and the codebase is designed for flexibility and extensibility. Overall, DigiRL aims to advance autonomous reinforcement learning for device control by providing a comprehensive framework, datasets, and pre-trained models to facilitate research and development in this domain.

https://github.com/DigiRL-agent/digirl

androidandroid-in-the-wildautonomous-reinforcement-learningautouicheckpoint-resumingcogagentdatasetsdevice-control-agentsdigirl-algorithmfiltered-behavior-cloningmulti-gpu-trainingmulti-machine-emulationoffline-to-online-trainingoffline-trainingonline-trainingpre-trained-modelsreinforcement-learningresearch-codetrajectory-video-recording