Endpoint trajectories

Added in 0.3.2a1: install with pip install --pre 'qenvs==0.3.2a1'.

rollout_metrics evaluates physical control amplitudes at right endpoints. These are privileged simulator diagnostics, not policy observations. An endpoint maximum is not a continuous-time maximum or an executable stopping policy.

Endpoint diagnostics for physical piecewise-constant controls.

This low-level helper does not apply environment action bounds or observations. Metrics are privileged simulator diagnostics, not policy inputs. Only right endpoints are sampled; their maximum is not a continuous-time maximum.

rollout_metrics(model, actions, *, dt, metrics, method)

Evaluate scalar metrics at the end of each physical control interval.

Parameters:

Name	Type	Description	Default
model	Model	a concrete model constructed outside JAX transformations.	required
actions	Array	real floating array (steps, model.n_controls), in physical units. No clipping, normalization or precision conversion is applied.	required
dt	float	positive finite control period, fixed outside jit.	required
metrics	Mapping[str, Callable[[Model, QArray], Array]]	nonempty mapping of names to pure (model, state) scalar-real callbacks. Capture this mapping outside jit, like dt and method.	required
method	Method	explicit dynamiqs solver, including its tolerances.	required

Returns:

Type	Description
Array	Times dt through steps*dt and a dict of metric arrays, each shape (steps,).
dict[str, Array]	No initial sample or state history is returned. Differentiation with
tuple[Array, dict[str, Array]]	respect to actions, jit and vmap are supported.

Raises:

Type	Description
ValueError	invalid shape, dtype, fixed dt or metric outputs.

Finite action and output checks belong at the host evaluation boundary; this scan does not convert traced arrays to NumPy or hide nonfinite scores.