Skip to content

Latest commit

 

History

History
46 lines (34 loc) · 3.78 KB

task_setups.md

File metadata and controls

46 lines (34 loc) · 3.78 KB

How to Set Up Tasks in MineDojo

We have predefined the task settings mentioned in the paper in the ./envs/tasks/task_specs.yaml file. Each task includes configurations for three stages: data collection, world model and behavior learning, and testing. For example, the task Harvest log in plains is defined as collect_rollouts_harvest_log_in_plains, harvest_log_in_plains, and test_harvest_log_in_plains for these three stages, respectively.

This document introduces the meaning and configuration guidelines for each field in the task settings, enabling you to define your custom tasks by following these examples.

Field Descriptions and Guidelines

  • task_id: Generally set to harvest. For more details, refer to Task Customization.

  • sim: Typically set to minedojo.

  • fast_reset: For more information, refer to Reset Mode.

    • During world model and behavior learning, fast_reset is often set to 5 or 10 to stabilize the training process.
    • During data collection and testing, fast_reset is usually set to 0 to collect more diverse data.

  • sim_specs: Details of the environment setup. See Task Customization for more information.

    • target_names: Specifies the names of the target items for harvest tasks.
    • target_quantities: Specifies the quantities of the target items for harvest tasks.
    • specified_biome: Defines the initial biome of the environment. See Specified Biome.
    • break_speed_multiplier: Adjusts the breaking speed of attacks. See Set Breaking Speed. Commonly set to 100 to simplify training.
    • initial_inventory: Specifies the initial inventory setup. See Initial Inventory.
    • initial_mobs: Specifies the initial mobs in the environment. See Spawn Mobs.

  • clip_specs: Configuration for using MineCLIP to calculate intrinsic rewards.

    • prompts: Short textual descriptions of the task, specified as list[str].

  • concentration_specs: Configuration for calculating intrinsic rewards based on affordance maps and jumping flags.

    • prompts: Short textual descriptions of the task, specified as list[str].
    • unet_checkpoint_dir: Path to the multimodal U-Net weight files.

  • reward_specs: Specifies rewards for collecting different items.

  • success_specs: Configures additional success criteria and actions upon success.

    • reward: Additional reward upon success.
    • any: Specifies that the task is considered successful if any of the listed conditions are met.
    • all: Specifies that the task is considered successful if all listed conditions are met.

  • terminal_specs: Configures conditions for ending an episode.

    • max_steps: Specifies the maximum number of steps per episode.

  • screenshot_specs: Configuration for saving observation images.

    • reset_flag: Whether to save images during reset. (Images during reset are often darker.)
    • step_flag: Whether to save images during each step.
    • save_freq: Interval between saved images on average.

  • LS_Imagine_specs: Settings for integrating the LS-Imagine algorithm with the MineDojo environment.

    • repeat: Number of times a single action is repeated. Typically set to 1.