A stand-in world for developing the runtime, driven by the actions it is given.
Monty's world is a Python process; until that bridge runs here
(docs/simulator-port.md), the runtime needs something deterministic to step.
This one moves a single agent according to the actions the motor system chose —
see Monty.Simulators.Synthetic.Kinematics for which actions it honours and
the conventions it follows — and answers the two payloads Monty's
Interface.step/1 returns: observations per agent and sensor, and the
proprioceptive state in global coordinates.
It has no geometry: nothing is out there, so an observation is what the sensor's own pose says, and the only feature is a value that follows the agent's x. That is enough for the runtime and for the pages that watch it, and it is worth being plain about what it is not: a run against this world says nothing about a learned object, while a run against a simulator says everything.
It is a simulator, not a ported module: nothing here claims to mirror a Python file.
How the runtime addresses this environment's parts in a module spec.
A world looking at scene, each sensor through its own camera.
What the world rendered on its last step, before any transform ran.
Nothing: this world places no object it could name.
Observations and proprioceptive state after actions.
@type t() :: %Monty.Simulators.Synthetic.Environment{ agent: Monty.Simulators.Synthetic.Kinematics.pose() | nil, last_observations: map() | nil, scene: Monty.Simulators.Synthetic.Scene.t() | nil, sensor_ids: [Monty.Simulators.Environment.sensor_id()], sensors: %{ optional(Monty.Simulators.Environment.sensor_id()) => Monty.Simulators.Synthetic.Kinematics.pose() }, views: %{optional(Monty.Simulators.Environment.sensor_id()) => keyword()}, world: Monty.Simulators.Synthetic.Kinematics.t() }
@spec agent_id() :: String.t()
How the runtime addresses this environment's parts in a module spec.
@spec new( [Monty.Simulators.Environment.sensor_id()], Monty.Simulators.Synthetic.Scene.t() | nil, keyword() ) :: t()
A world looking at scene, each sensor through its own camera.
:views is a camera per sensor, keyed by sensor id, in the terms
Monty.Simulators.Synthetic.Scene.render/3 takes — a resolution, a zoom and a
field of view. A sensor with no camera renders one ray down its own axis.
:agent and :sensors are where the episode starts, as an experiment's
simulator is configured.
Geometry is a choice, not an assumption: with a scene the world renders depth, colour and semantic id for every sensor, which is what a ported surface policy steers by; without one it reports only where the sensors are, and a percept built from it says its pose is undefined.
What the world rendered on its last step, before any transform ran.
A page that shows the object wants the picture the world drew, not the one a
transform rewrote: MissingToMaxDepth turns "no return" into a metre, and a
surface experiment clips everything past five centimetres, which is right for
matching and useless for looking at.
@spec primary_target(t()) :: nil
Nothing: this world places no object it could name.
Monty's worlds sample an object and say which; the stand-in has one sphere at the centre an experiment's scene put it, and calling that an object would tell a model to score itself against something nobody chose.
@spec step(t(), [Monty.Frameworks.Actions.Action.t()]) :: {map(), map(), t()}
Observations and proprioceptive state after actions.
Keys and units follow Monty: metres, sensor positions in the world, the agent's pose absolute, each sensor's pose relative to the agent.