NEWS!!

• (2024-07-14 v4.0)
  • Added new action and observation type of ADB
  • Enabled input of common UTF-8 strings for TEXT action
  • Enabled fuzzy match method for screen text events. Enabled triggering threshold for fuzzy match modes.
  • Migrated from dm_env to android_env.interfaces to distinguish successful, failed, and truncated episode ends. Updated episode end events to control its triggering in the task definition file more conviniently.
  • Added cache_until field for event slots to correctly trigger an AND node whose sub-events are expected to be triggered simultaneously. Now an activatated event can be cached temporarily until another triggered event clears it.
  • Added null_listener for target-less event nodes.
  • Applied image compression in RemoteSimulator.
  • Migrated from gym to gymnasium.
  • Updates to VhIoWrapper and TapActionWrapper
  • Minor updates to annotation tool.

See our Change Log for details. The documents will be revised soon. A new tutorial w.r.t. episode event management is on plan.

• (2024-04-30 v3.6)

  • Updated function to load a remote simulator to enable providing the remote resources with a different path with the path of the local task definition file.
  • Updated task template toolkit, added new slot modifiers and sytaxes for task config file.
  • Fixed known bugs.

• (2023-12-18 v3.5)

  • Owing to the long time delay of VH check and screenshot check, we updated the mechanism of managing the check time. By this way, the requirement of sufficient check for the episode events and the resulted long delay can be balanced.
  • Added multiple rating methods to ResponseEvent: regex matching, fuzzy matching, and vector encoding matching.
  • Improved VhIoWrapper and TapActionWrapper. Added support to SCROLL and TYPE to TapActionWrapper.
  • Optimized RemoteSimulator. In order to reduce the delay of network transfering, enabled action batch to send and execute a group of actions and enabled resizing the image before and after transferring to shrink the transferred data.
  • Merged annotation tool to the main branch. The original annotation-tool branch is deprecated.
  • Added support to ResponseEvent to annotation tool.
  • Supplemented several commandline options to annotation-tool.

For more details, please see our Change Log and Documents.

• (2023-10-31 v3.0) Migrated VH node specification from the original VH path to Mobile-Env-customized CSS selector (me-selector) and added repeatability control to EventSlots. Repeatability control for EventSlots may be useful to prevent repetitive triggering of an OR-type virtual event combining multiple types of event sources.

Please see our Change Log and Document.

• (2023-09-21 v2.1) Added REMOTE SIMULATOR to solve the problem that hardware-based acceleration for virtualization is not enabled on many GPU clusters

Please see our Change Log and Document.

• (2023-06-30 v2.0) New type of event "response to human user" (RHU, ResponseEvent). Now enables the agent to generate response to human user and parses episode signales from it. This will enable interaction tasks like question-answering, retrieval, etc.

Please see our Change Log, Usage Document, and Task Definition Document.

Mobile-Env: Building Qualified Evaluation Benchmarks for GUI Interaction

Mobile-Env is a interaction platform for building evaluation benchmarks for GUI interaction and evaluating and training GUI agents. Our paper is available at arXiv.

Mobile-Env is developed based on AndroidEnv. The agent can take the screenshot and the view hierarchy (disabled defaultly for the long latency) as the observation and take a touch or type a token as the action to interact with the Android apps. Several episode signals like step instructions, rewards, or the episode end will be informed during interaction at some crucial steps. A so-called crucial step may be opening a target page, srolling to a correct area, etc. and is depending on the specific task definition.

The proposed WikiHow task set is available at the Hugging Face Platform.

• Documents in Chinese (中文文档)

Index

• Evaluating and Training Agents on Mobile-Env
• Extending a New Environment (App) or a New Task Based on Mobile-Env
• Certificate Pinning Problem & Solutions
• Miscellaneous Auxiliary Tools

Platform Features

Mobile-Env is a flexible, adaptable, and easily-extendable platform for InfoUI interaction with the following features:

• Both screenshot and view hierarchy are provided as the observation. The touch and token typing are provided as the action. Wrappers are also supported to customize the observation and action spaces. Thus, both visual-based and text-based agents, both agents with continuous action space and discrete action space, can be evaluated on Mobile-Env.
• New tasks can be easily extended through task definition files.
• Multiple sources are enabled to parse the task events from the operating system: screen text, screen icon, view hierarchy, and the system log, which makes Mobile-Env capable of adapting to most real-world apps without dedicated development. (Screen text and screen icon will be enabled with an external OCR tool and icon recognition tool. Currently, a wrapper of EasyOCR is integrated in the platform and can be enabled directly. An intergrated icon model will be embedded soon as well.)

Getting Started

Installation

Install from PyPI:

pip insall mobile-env-rl

or clone the repository and build locally.

git clone https://github.com/X-LANCE/Mobile-Env
cd Mobile-Env
pip install .

Several Docker images with well-configured Android AVD are also available.

Load and Run Mobile-Env for Evaluation or Training

Before loading the Mobile-Env environment, you will need to set up an Android Emulator device. Then you can load the environment with some existing task definitions and start your experiments. A detailed guidance is provided in Evaluating and Traning Agents on Mobile-Env. Several examples with a random agent or a human agent is also provided under examples.

Extend a New Environment or a New Task

To extend a new environment for Mobile-Env, the environment designer needs to prepare the app package and ensure that the package manages to launch and run on some versions of Android Emulator. If the app requires varying online data, the necessary data should be crawled and dumped and then be replayed for a consistent evaluation. In such case, the designer is supposed to validate the certain effectiveness of certificate unpinning plan for the package. As regards to extend new tasks, task definition files are just required. Detailed instructions can be found in Extending a New Environment (App) or a New Task Based on Mobile-Env.

Several demo task definitions are provided under demos. Three of them are migrated from AndroidEnv:

• classic_2048.m.textproto - Classic 2048 game.
• accessibility_forwarder_clock_set_timer.m.textproto - A simple task requiring the agent to reset a running timer.
• systemui_egg_land_default.m.textproto - Flappy Droid. An open-sourced implementation of classic game, Flappy Bird.

Another one, openmoneybox.add_billings.textproto is defined upon an open-sourced billing app, OpenMoneyBox. Details are referred to in the task definition files.

Miscellaneous Auxiliary Tools

We also developed an annotation tool for the human demonstrations, and a suite of template tool to auto-generate task definitions according to templates and to combine multiple task definitions to form a multi-step task. The details are referred to in Miscellaneous Auxiliary Tools.

Reference Time-Consuming and Memory Usage o Mobile-Env

The data are measured under the configuration below:

• OS and hardware:
  • Operating System: Manjaro 23.1.0 Vulcan
  • Kernel Version: x86_64 Linux 6.1.64-1-MANJARO
  • CPU: Intel Core i7-10700 @ 16x 4.8GHz
  • GPU: NVIDIA GeForce RTX 3090
  • RAM: 64 GB
  • KVM acceleration enabled
• Android development tools
  • Android emulator version 32.1.14.0
  • Android platform tools 34.0.4
  • libvert 1:9.9.0
• Python & packages
  • Python 3.8.16
  • EasyOCR 1.7.2
  • sentence-transformers 2.2.2
• Android Virtual Device
  • Device type: Pixel 2
  • API version: API 30
  • OS Variant: Google APIs
  • CPU cores: 4
  • Memory: 8 GB
  • Screen size: 1080×1920

Item	Avg Time	Time Std Dev
TOUCH action	410.50 µs	64.71 µs
LIFT action	412.30 µs	84.18 µs
TEXT action	1.30 s 0.58s	0.28 s 0.03 s
screenshot capturing	19.94 ms	21.47 ms
invocation of Sentence Transformer（all-MiniLM-L12-v2）	8.51 ms	0.17 ms
VH capturing	2.53 s	1.90 s
invocation of EasyOCR	0.44 s	0.08 s

When only an app of WikiHow 2.9.6 is running, the Android emulator occupies 6,031 MiB of virtual memory and 3,444 MiB of residual memory.

About

This library is developed and maintained by SJTU X-Lance. The corresponding paper is available at https://arxiv.org/abs/2305.08144.

If you find Mobile-Env useful in your research, you can cite the project using the following BibTeX:

@article{DanyangZhang2023_MobileEnv,
  title     = {{Mobile-Env}: Building Qualified Evaluation Benchmarks for LLM-GUI Interaction},
  author    = {Danyang Zhang and
               Zhennan Shen and
               Rui Xie and
               Situo Zhang and
               Tianbao Xie and
               Zihan Zhao and
               Siyuan Chen and
               Lu Chen and
               Hongshen Xu and
               Ruisheng Cao and
               Kai Yu},
  journal   = {CoRR},
  volume    = {abs/2305.08144},
  year      = {2023},
  url       = {https://arxiv.org/abs/2305.08144},
  eprinttype = {arXiv},
  eprint    = {2305.08144},
}