Skip to content
This repository was archived by the owner on Jun 8, 2024. It is now read-only.

index.md

Latest commit

 

History

History
80 lines (62 loc) · 3.42 KB

index.md

File metadata and controls

80 lines (62 loc) · 3.42 KB
Raw

Tasks

Alchemy consists of a single procedurally generated task / level, called as:

'alchemy/perceptual_mapping_randomized_with_rotation_and_random_bottleneck' -- stone perceptual features can be rotated 45 degrees with respect to the latent feature axes. Only 2 axes can be off-axis at a time (e.g. x and y are rotated 45 degrees, but z is axis-aligned).

Actions

The 3D environment provides the following actions:

  • MOVE_BACK_FORWARD: range [-1.0, 1.0]
  • STRAFE_LEFT_RIGHT: range [-1.0, 1.0]
  • LOOK_DOWN_UP: range [-1.0, 1.0]
  • LOOK_LEFT_RIGHT: range [-1.0, 1.0]
  • HAND_ROTATE_AROUND_FORWARD: range [-1.0, 1.0]
  • HAND_ROTATE_AROUND_RIGHT: range [-1.0, 1.0]
  • HAND_ROTATE_AROUND_UP: range [-1.0, 1.0]
  • HAND_PUSH_PULL: range [-10.0, 10.0]
  • HAND_GRIP: range [0.0, 1.0]

Each action is a double scalar. It is not compulsory to send a value for each action every step, but note that actions are "sticky", meaning an action's value will only change when a new value is provided. For example:

env = dm_alchemy.load_from_docker(settings)
env.reset()
env.step({'STRAFE_LEFT_RIGHT': -1.0}) # Result: strafe Left.
env.step({'MOVE_BACK_FORWARD': 1.0}) # Result: strafe left & move backward.

env.step({'STRAFE_LEFT_RIGHT': 0.0,
          'MOVE_BACK_FORWARD': 0.0}) # Result: stationary.

The symbolic version of Alchemy takes an integer action, corresponding to either performing a no-op, ending the trial (optional, specified with a flag), or putting stone i into potion j or the cauldron (see symbolic_alchemy.py for more details).

Observations

The 3D environment provides the following observations:

  • RGB_INTERLEAVED: First person RGB camera observation. The width and height can be adjusted through the EnvironmentSettings, but the observation will always have a fixed 4:3 aspect ratio.
  • ACCELERATION: The avatar's "acceleration" vector, in meters/second^2. The components are measured, in the (right, up, forward) directions relative to the direction the avatar is facing. This is represented as a vector of 3 floats.
  • HAND_FORCE: The force, as a fraction of the maximum grip strength, the hand is applying to the object it's manipulating. If holding an object, this represents how close to dropping it the avatar is. This is represented as a float with a value between 0 and 1.
  • HAND_IS_HOLDING: Set to 1 if the avatar is holding an object, 0 otherwise.
  • HAND_DISTANCE: Distance of hand from face, expressed as a fraction of min and max distance. 0 when not holding an object, otherwise between 0 and 1.
  • Score: The agent's cumulative score.

The symbolic version of Alchemy returns a binary or discrete vector representation (normalized to [-1.0, 1.0]) of the existing stones and potions, in terms of their features, as well as optional flags for whether or not they're used (see symbolic_alchemy.py for more details).

Configurable environment settings

Required attributes:

  • seed: Seed to initialize the environment's random number generator.
  • level_name: Name of the level to load as they appear in the first section.

Optional attributes:

  • width: Width (in pixels) of the desired RGB observation; defaults to 96.
  • height: Height (in pixels) of the desired RGB observation; defaults to 72.
  • num_action_repeats: Number of times to step the environment with the provided action in calls to step(); defaults to 1.