Scene Generation
The “output” of a Scenic program has two parts: a scene describing a configuration of physical objects, and a policy defining how those objects behave over time. The latter is relevant only for running dynamic simulations from a Scenic program, and is discussed in our page on Execution of Dynamic Scenarios. In this page, we describe how scenes are generated from a Scenic program.
In Scenic, a scene consists of the following data:
a set of objects present in the scene (one of which may be designated the
ego
object);concrete values for all of the properties of these objects, such as
position
,heading
, etc.;concrete values for each global parameter.
A Scenic program defines a probability distribution over such scenes in the usual way for imperative probabilistic programming languages with constraints (often called observations).
Running the program ignoring any require
statements and making random choices whenever a distribution is evaluated yields a distribution over possible executions of the program and therefore over generated scenes.
Then any executions which violate a require
condition are discarded, normalizing the probabilities of the remaining executions.
The Scenic tool samples from this distribution using rejection sampling: repeatedly sampling scenes until one is found which satisfies the requirements.
This approach has the advantage of allowing arbitrarily-complex requirements and sampling from the exact distribution we want.
However, if the requirements have a low probability of being satisfied, it may take many iterations to find a valid scene: in the worst case, if the requirements cannot be satisfied, rejection sampling will run forever (although the Scenario.generate
function imposes a finite limit on the number of iterations by default).
To reduce the number of iterations required in some common cases, Scenic applies several “pruning” techniques to exclude parts of the scene space which violate the requirements ahead of time (this is done during compilation; see our paper for details).
The scene generation procedure then works as follows:
Decide which user-defined requirements will be enforced for this sample (soft requirements have only some probability of being required).
Invoke the external sampler to sample any external parameters.
Sample values for all distributions defined in the scene (all expressions which have random values, represented internally as
Distribution
objects).Check if the sampled values satisfy the built-in and user-defined requirements: if not, reject the sample and repeat from step (2).