Monty.Simulators.Synthetic.Kinematics (ExTbpMonty.Core v0.1.0)

Copy Markdown View Source

Where the stand-in world puts the agent after an action.

Monty's world is MuJoCo, and its actuator defines the conventions this module follows, because a policy that steers one must steer the other:

  • MoveForward moves along the agent's local −Z — towards whatever the sensor is looking at — and a negative distance moves it backwards;
  • TurnLeft / TurnRight yaw the agent about its local Y, positive to the left;
  • LookUp / LookDown pitch the sensors about their local X, and the agent stays where it points;
  • MoveTangentially moves along the action's direction, which is already a world vector: the surface policies rotate their heading by the agent's rotation before they answer, so it must not be rotated again;
  • OrientVertical / OrientHorizontal orbit the agent about its own axes: a side step, then a pitch or yaw, then a step along the agent's −Z. Monty's actuator "ASSUMES" the distances were calculated to orbit the object, and takes the surface into account not at all — so neither does this.

What is not honoured: the actions that set a pose outright (SetAgentPose, SetSensorPose, SetYaw, …). Asking for one is an error rather than a silent no-op — a world that quietly ignores an action makes an agent look broken in a way that the log cannot explain.

Summary

Types

pose()

Where one body sits: a position in the world, and a rotation from its own frame to the world's.

t()

Functions

new(sensor_ids, options \\ [])

A world with the agent where the experiment starts it.

sensor_poses(world)

Each sensor's own pose, relative to the agent, as proprioception reports it.

sensor_positions(world)

Where each sensor is, as the environment reports it.

sensor_world_poses(world)

Where each sensor is and which way it faces, in the world.

step(world, actions)

The world after every action of one step, in the order they were given.

Types

pose()

@type pose() :: %{
  position: Monty.Geometry.Rotation.vector(),
  rotation: Monty.Geometry.Rotation.t()
}

Where one body sits: a position in the world, and a rotation from its own frame to the world's.

t()

@type t() :: %{
  agent: pose(),
  sensors: %{optional(Monty.Simulators.Environment.sensor_id()) => pose()}
}

Functions

new(sensor_ids, options \\ [])

@spec new([Monty.Simulators.Environment.sensor_id()], keyword()) :: t()

A world with the agent where the experiment starts it.

Options: :agent, the agent's start pose, and :sensors, a map of sensor id to that sensor's pose relative to it — both as an experiment's simulator is configured. A sensor the options do not place rides a decimetre back from the agent and a centimetre apart from the others, which is the stand-in's own geometry: a world with no scene description still needs two cameras to see two places.

sensor_poses(world)

@spec sensor_poses(t()) :: %{
  optional(Monty.Simulators.Environment.sensor_id()) => pose()
}

Each sensor's own pose, relative to the agent, as proprioception reports it.

sensor_positions(world)

@spec sensor_positions(t()) :: %{
  optional(Monty.Simulators.Environment.sensor_id()) =>
    Monty.Geometry.Rotation.vector()
}

Where each sensor is, as the environment reports it.

sensor_world_poses(world)

@spec sensor_world_poses(t()) :: %{
  optional(Monty.Simulators.Environment.sensor_id()) => pose()
}

Where each sensor is and which way it faces, in the world.

The sensor's own frame is the agent's turned by however the sensor is mounted on it, so raising the agent's yaw carries its cameras round too.

step(world, actions)

@spec step(t(), [Monty.Frameworks.Actions.Action.t()]) :: t()

The world after every action of one step, in the order they were given.