Create a Task Preset

Start from FirstTask in your own package. A preset assembles existing components into an explicit interaction.

In the first task, the tool presses a button and the stage's goal checks that it is depressed. For an action order, record the relevant tool calls and verify their logs. For a judged answer, add a text-only stage. These choices determine what “success” means; stages and interaction explains them together.

Minimum declaration

Define maniskill_env_id, Tools_cls, situation_init, and a nonempty stages list. Call super().__init__() to obtain default metadata, initialization parameters, name, and tool constants.

Create mutable attributes and stage lists per instance. A class-level SituationInit mutated by one preset instance can unintentionally affect another.

Parameterize the interaction

The first tutorial accepts button and passes it to its stage. Keep the instruction and verification goal derived from the same argument. Extra test-tools CLI arguments are passed to your preset constructor.

Customize initialization

Inside your preset constructor, after super().__init__():

self.initialization_parameters.env_options = {"use_visual_assets": False}

Only use options supported by your chosen environment. For multiple robots, configure agent_names in InitializationParameters and keep their order aligned with the environment.

SituationInit.attributes supplies initial vocabulary; all_task_attributes supplies future possible values when vocabulary changes. Initial memory is a dictionary. For example, the provided history agents accept memory={"memory_list": ["Keep fragile items separate."]} for initial rules. This is a consumer convention, not a universal scenario memory schema. The agent package manages its subsequent memory and history.

Register and test

Add the class reference to your package's manifest presets. Run magma-scenarios show my_buttons, then magma-scenarios test-tools my_buttons.FirstTask --nb-env 1.

A preset is useful before procedural requests exist. You can debug tools and action goals without running a data-generation pipeline.

Reference: task fields and validation.