mujoco_worker¶

Base class for parallel MuJoCo robot evaluation.

Users subclass MuJoCoWorkerBase and implement evaluate() with their experiment-specific fitness logic. The instance is passed directly to multiprocessing.Pool.imap — it is picklable because it is a named class.

Module Contents¶

Classes¶

EvalConfig

Minimal config passed to each worker process.

MuJoCoWorkerBase

Callable base class for MuJoCo robot evaluation in worker processes.
class EvalConfig¶

Minimal config passed to each worker process.

Users may subclass this to carry additional experiment-specific fields.

spawn_position: tuple[float, float, float]¶
target_position: tuple[float, float, float]¶
seed: int | None = None¶
class MuJoCoWorkerBase¶

Bases: abc.ABC

digraph inheritance1cc44eae13 { bgcolor=transparent; rankdir=LR; size="8.0, 12.0"; "ABC" [fillcolor=white,fontname="Vera Sans, DejaVu Sans, Liberation Sans, Arial, Helvetica, sans",fontsize=10,height=0.25,shape=box,style="setlinewidth(0.5),filled",tooltip="Helper class that provides a standard way to create an ABC using"]; "MuJoCoWorkerBase" [URL="#ariel.simulation.mujoco_worker.MuJoCoWorkerBase",fillcolor=white,fontname="Vera Sans, DejaVu Sans, Liberation Sans, Arial, Helvetica, sans",fontsize=10,height=0.25,shape=box,style="setlinewidth(0.5),filled",target="_top",tooltip="Callable base class for MuJoCo robot evaluation in worker processes."]; "ABC" -> "MuJoCoWorkerBase" [arrowsize=0.5,style="setlinewidth(0.5)"]; }

Callable base class for MuJoCo robot evaluation in worker processes.

Subclass and implement evaluate(). Pass an instance to multiprocessing.Pool.imap as the map function:

worker = MyWorker()
with pool:
    results = list(pool.imap(worker, eval_args))

where eval_args is a list of (xml_string, EvalConfig) tuples.

__call__(args: tuple[str, EvalConfig]) → float¶

Entry point called by the worker process.

Loads the MuJoCo model from XML and delegates to evaluate(). Returns the result of evaluate(), or raises if the XML is invalid.

Returns:

The fitness value returned by evaluate().

Return type:

float

abstract evaluate(model: mujoco.MjModel, data: mujoco.MjData, config: EvalConfig) → float¶

Evaluate a robot and return a scalar fitness value.

Called once per individual per generation inside a worker process. PyTorch is already limited to 1 thread when this is called.

Parameters:

  • model (Loaded MuJoCo model.)

  • data (Corresponding MjData, ready for simulation.)

  • config (Experiment configuration passed from the main process.)

Returns:

  • A scalar fitness value (convention (lower = better, but the)

  • sign is entirely up to the user's experiment).