Skip to content

Latest commit

 

History

History
38 lines (34 loc) · 3.75 KB

action_space.md

File metadata and controls

38 lines (34 loc) · 3.75 KB
Raw

Action space

HAZARD support two different types of action set. Users can choose one of them when implementing agent algorithms.

In this page, we introduce all supported actions in the format of <action_name> (<parameter>): <action description>

High-level action set

  • walk_to (obj_id): Agent will try to walk to the
  • pick_up (obj_id): Agent will pick up the object with given obj_id. If obj_id is None, agent will try to pick up the object it just walked to.
  • drop (obj_id): Agent will drop the object with given obj_id. In fire or flood scenario, agent will simply put the object in its bag, while in wind scenario it will drop the object to a shopping cart or to the ground. If obj_id is None, agent will try to drop the object it holding.
  • explore: Agent will simply look round without changing the location of itself.

Low-level action set

(Adopted from TDW replicant)

Low-level actions can be used in format of ('low_level.<action_name>', <params_dict>). Possible <action_name>s are listed below:

  • move_by:
    • param distance: The target distance. If less than 0, the Replicant will walk backwards.
    • param reset_arms: If True, reset the arms to their neutral positions while beginning the walk cycle.
    • param reset_arms_duration: The speed at which the arms are reset in seconds.
    • param scale_reset_arms_duration: If True, reset_arms_duration will be multiplied by framerate / 60), ensuring smoother motions at faster-than-life simulation speeds.
  • turn_by:
    • param angle: The target angle in degrees. Positive value = clockwise turn.
  • turn_to:
    • param target: The target. If dict: A position as an x, y, z dictionary. If numpy array: A position as an [x, y, z] numpy array.
  • reach_for:
    • param target: The target(s). This can be a list (one target per hand) or a single value (the hand's target). If int: An object ID. If dict: A position as an x, y, z dictionary. If numpy array: A position as an [x, y, z] numpy array.
    • param arm: The Arm value(s) that will reach for each target as a single value or a list. Example: Arm.left or [Arm.left, Arm.right].
    • param absolute: If True, the target position is in world space coordinates. If False, the target position is relative to the Replicant. Ignored if target is an int.
    • param offhand_follows: If True, the offhand will follow the primary hand, meaning that it will maintain the same relative position. Ignored if arm is a list or target is an int.
    • param from_held: If False, the Replicant will try to move its hand to the target. If True, the Replicant will try to move its held object to the target. This is ignored if the hand isn't holding an object.
    • param held_point: The bounds point of the held object from which the offset will be calculated. Can be "bottom", "top", etc. For example, if this is "bottom", the Replicant will move the bottom point of its held object to the target. This is ignored if from_held == False or ths hand isn't holding an object.

Action execution result

  • success: action executed successfully
  • fail: fail to execute the action with a reason, typically the following ones
    • max steps reached: walk_to will fail if the navigation algorithm reaches its max step. Typically it happens when the agent is too far away from the target object or it can not find a path to the target object.
    • target not in vision or memory: walk_to will fail if the given obj_id is not in agent's vision or memory.
    • cannot grasp, maybe too far: pick_up will fail when the target object is too far from the agent.
    • not holding an object: drop will fail if the agent is not holding any objects.