Skip to content

Latest commit

 

History

History
107 lines (76 loc) · 5.93 KB

README.md

File metadata and controls

107 lines (76 loc) · 5.93 KB
Raw

Dataset

This section contains the description of the dataset used for Synthesizing Environment-Aware Activities via Activity Sketches. Note that you need to use version v1.0.0 of the simulator to use this dataset.

Move the dataset of programs and environments representing activities in this folder. The dataset should have the following structure:

programs_processed_precond_nograb_morepreconds
|── initstate
├── withoutconds
├── executable_programs
|   ├── TrimmedTestScene7_graph
|	└── ...
└── state_list
	├── TrimmedTestScene7_graph
   	└── ...

Following, we describe the meaning of each folder, but first, let's introduce VirtualHome 2 main concepts, Environments and Programs.

Environments

VirtualHome is composed of 7 scenes where activities can be executed. Each scene is encoded in a .json file containing a node for every object and edges between them representing relationships. Each environment can be updated by modifying the corresponding .json file.

You can generate a graph for a given apartment by using unity simulator.

scene_id = 0 # Scenes go from 0 - 6
comm = UnityCommunication()
comm.reset(scene_id)
s, graph = comm.environment_graph()

You can check in the demo examples on how to read and update graphs.

You can view the supported objects and states in Resources.

Programs

Activities in VirtualHome are executed through programs. Each program is a sequence of instructions representing atomic actions that the agent will execute. Each program is stored in a .txt file with the following format.

Program title
Program description


[ACTION_NAME] arg1 arg2
[ACTION_NAME] arg1 arg2
...

Each each argument has the format <OBJECT_NAME> (ID_OBJECT). The number of arguments depends on the action type. The (ID_OBJECT) is used to tell the simulator that the actions should be done on the same object instance. For example a program as

[Walk] <glass> (1)
[Grab] <glass> (1)

Indicates that the agent should first walk to a glass, and then grab that same glass.

The programs can also have a precondition file, specifying the state of the objects before the program is executed.

You can view the supported actions, objects and preconditions in Resources.

Dataset

The dataset is structured in 5 folders: initstate, withoutconds, executable_programs, init_and_final_graph and state_list. We describe this structure below.

withoutconds

This folder contains the originally collected programs, following the format specified above.

initstate

We previously mentioned that programs can have preconditions, specifying certain constraints that an environment should satisfy in order to execute the program. For instance, if a program starts with [Open] <fridge> (1), it is clear that <fridge> (1) should be initally closed.

We did not collect such preconditions directly, but use a set of rules to extract them for every program, using ../dataset_utils/add_preconds.py. For each program in without conds, you can find a json with the same program path in initstate, containing the preconds.

executable_programs

So far we have talked about programs and environments separately, but we want to execute each program in the environments. Let's say that we want to execute the previously mentioned program

[Walk] <glass> (1)
[Grab] <glass> (1)

In an environment that has 3 glasses with ids 23, 45, 58. It is clear that the glass we walk to and grab should be the same, and that if there was a <glass> (2) we should be using a different glass, but how do we know if we should use 23, 45, 58? This is what the executable_programs folder is for.

For each program and its preconditions, we select one of the TrimmedGraphs representing an environment and use graph_evolve simulator to initialize the environment following the preconditions. Finally, we assign the objects in the program to objects in the graph. If the simulator decides that <glass> (1) should correspond to the glass with id 23, a file will be created in executable_programs/TrimmedTestScene{graph_id}_graph looking as follows.

[Walk] <glass> (1.23)
[Grab] <glass> (1.23)

You can check demo/generate_snapshot.py to see an example of how a program is created.

Note: When executing programs in the unity_simulator, you will normally specify the id corresponding to an object existing in the environment, but you can also let the simulator decide which object (for instance which of the glasses to use) by setting the flag find_solution=True.

init_and_final_graph, state_list

We have now modified the program to be executable in a given environment, and also modified the environment to match the programs preconditions. How does the environment look now? The graph is found in these 2 folders.

init_and_final_graph contains the first graph (before executing the program) and the last graph, when the program is executed using graph_evolve simulator. If you want to execute the program in the unity_simulator, make sure to set the environment according to id. That is:

comm.reset(graph_id - 1)
with open('init_and_final_graph/{}'.format(graph_file), 'r') as f:
	graphs = json.load(f)
	first_graph = graphs['init_graph']
comm.expand_scene(first_graph)

TLDR;

Knowing the structure of these folders is useful to understand and use the dataset, but you may get a faster glance by looking at ../demo/generate_snapshot.py. Additionally, you can find some utility functions using the dataset in ../dataset_utils/execute_script_utils.py.