Implements tasks and measurements needed for training and benchmarking of
- class Action
- An action that can be performed by an agent solving a task in environment.
For example for navigation task action classes will be:
MoveForwardAction, TurnLeftAction, TurnRightAction. The action can use
Taskmembers to pass a state to another action, as well as keep own state and reset when new episode starts.
- class ActionSpace
- A dictionary of
EmbodiedTaskactions and their argument spaces.
- class EmbodiedTask
- Base class for embodied tasks.
EmbodiedTaskholds definition of a task that agent needs to solve: action space, observation space, measures, simulator usage.
EmbodiedTaskhas reset() and step() methods that are called by
EmbodiedTaskis the one of main dimensions for the framework extension. Once new embodied task is introduced implementation of
EmbodiedTaskis a formal definition of the task that opens opportunity for others to propose solutions and include it into benchmark results.
- class EmptySpace
gym.Spacethat reflects arguments space for action that doesn’t have arguments. Needed for consistency ang always samples None value.
- class Measure
- Represents a measure that provides measurement on top of environment and task.
- class Measurements
- Represents a set of Measures, with each Measure being identified through a unique id.
- class Metrics
- Dictionary containing measurements.
- class SimulatorAction
EmbodiedTaskaction that is wrapping simulator action.