DroidStar is an active learning tool that synthesizes behavioral specifications for event-driven framework classes that explain how and when their callbacks occur.

View the Project on GitHub cuplv/droidstar


DroidStar is an active learning tool that synthesizes behavioral specifications for event-driven framework classes that explain how and when their callbacks occur.

In Android application programming, understanding when the application is allowed to call into the framework (callins) and when the framework may call back to the application (callbacks) is difficult. The object-oriented type of Android classes does not provide this extra “callback typestate” information, and writing manual tests for different sequences of asynchronous events is extremely tedious. DroidStar automatically chooses and executes test sequences in order to fully explore the possible states of an Android class, using the results of these tests to generate a specification of all possible callin and callback orderings. This specification can be used by the developer of the class to check that their implementation does what they intend, and it can serve as documentation to quickly explain to users of the class how to interact with it.

This repository holds a growing set of specification experiments which can be quickly run with a template Android application.

Quick start

Your environment will need: git, dot (a graph-drawing program provided by the graphviz package), and sbt.

Clone the repository and connect your Android device.

$ git clone https://github.com/cuplv/droidstar
$ sudo adb start-server
$ # Connect your device
$ adb devices

If adb devices lists your device, you are good to go.

Now give a command to sbt that will build, install, and run the experiment application on your device.

$ ./run-experiments

You should see an empty screen appear on your device. You can now follow the experiment’s progress.

$ ./track-progress

By default, the experiment application will learn the callback typestate for the AsyncTask class. The tracker will tell you when the experiment is complete (and your device will return to its home-screen). At this point you can fetch the results from the device.

$ ./fetch-results

This command will place the results of the experiment in a ./results directory. This directory will contain a .png image that shows the learned callback typestate for the AsyncTask class.

Interpreting the results

When you fetch the results of an experiment from the phone, you get back three files.

Running other experiments

An experiment tests a list of LearningPurpose instances in sequence, allowing you to fetch the results all at once at the end. You can edit MainActivity.scala, to change the classes you’d like to test by adding them to the purposes list.

The classes you can choose from for experiments are found in the lp source code directory (and also in the older java lp directory; they inmplement the same interface). If you would like to experiment on a new class, you can use these as examples to write a new LearningPurpose for it. Also, check out the LearningPurpose class source file for some useful comments on what its various methods and options are for.

Writing an experiment

In order to perform your own experiment on a class you are interested in, you must write an instance of the LearningPurpose abstract class that tells droidstar how to explore its behavior.

We begin with the boiler-plate imports. Add any imports you need here, and replace the name AsyncTaskLP with whatever your experiment should be called. By convention, we name them $(class under study)LP.

package edu.colorado.plv.droidstar
package experiments.lp

import android.content.Context
import android.os.Handler.Callback
import android.os.AsyncTask
import scala.collection.JavaConverters._

class AsyncTaskLP(c: Context) extends LearningPurpose(c) {

Now that we are defining our subclass, it is helpful to start by defining the various identifiers you will use up front, so that they are not mis-typed later. These String identifiers will be associated with code snippets and used to mark results in the automaton that is produced.

You will need one for each distinct input and output that you are studying.

  // inputs
  val execute = "exec"
  val cancel = "cancel"

  // outputs
  val cancelled = "on_cancelled"
  val postexec = "on_postexec"
  val preexec = "on_preexec"

Next, establish the mutable state that droidstar will work on. If the focus of your experiment is a singe class, such as the AsyncTask class in this example, this state will simply be an object of the class. The object does not need to be instantiated at this point; it will be re-initialized at the beginning of each testing round.

  var task: AsyncTask[AnyRef,AnyRef,AnyRef] = null

In most cases, you will need to extend the class you are studying in order to instrument its callbacks with reports that droidstar can see. Here, we define a simple AsyncTask instance that waits a little while as its task and reports callback identifiers (that we defined in the previous step) using the repsond() method that LearningPurpose provides.

  class SimpleTask(localCounter: Int) extends AsyncTask[AnyRef,AnyRef,AnyRef] {

    override def doInBackground(ss: AnyRef*): AnyRef = {
      try {Thread.sleep(200)}
      catch {
        case _ : Throwable => logl("Sleep problem?")

    override def onPostExecute(s: AnyRef): Unit = respond(postexec)

    override def onPreExecute(): Unit = respond(preexec)

We now define the steps droistar takes to set up a test. Usually this is a code snippet that initializes the mutable state we have established. Here we initialize the task variable we previously declared with an instance of SimpleTask, disabling and discarding any task left over from a previous test.

  override def resetActions(c: Context, b: Callback): String = {
    if (task != null) {
    task = new SimpleTask(0)

Now define the LearningPurpose.uniqueInputSet(), a list of String values, as the list of your input identifiers.

  override def uniqueInputSet(): java.util.List[String] =

This list will be used to generate test sequences. So that droistar can actually execute each test, you must associate each input identifier in the list with a code snippet that acts on the mutable state you have established.

  override def giveInput(i: String, altKey: Int): Unit = i match {
    case `execute` => task.execute("asdf")
    case `cancel` => task.cancel(false)

    case _ => {
      logl("Unknown command to AsyncTask")
      throw new IllegalArgumentException("Unknown command to AsyncTask")

Almost finished! All that remains is a handful of optional settings that you may need to adjust for your experiment to be useful. The few most important ones appear here; the full list of modifiable settings can be found in the LearningPurpose source file.

The betaTimeout is an integer representing the number of milliseconds droidstar should wait to receive a callback. This timeout is very important; it should be greater than the amount of time you think any callback you are tracking should take so that none are missed.

  override def betaTimeout(): Int = 500

The isError function takes an output identifier (as reported by a callback using the respond method) and states whether it should be considered an error. This is used to make an input as “not enabled” even if it didn’t throw a synchronous error. Some callins report that they failed via callback.

  override def isError(o: String): Boolean = false

We don’t have any of these for AsyncTask.

The last option is a name that droistar will use to title the results it produces. It is important to make sure if you are running several experiments together that each of their LearningPurposes has a different shortName value.

  override def shortName(): String = "AsyncTask"