1 rollouts.inspection.selector_math

rollouts.inspection.selector_math

Retained selector shadow normalization and acquisition-displacement heading cost.

1.1 Functions

Name	Description
horizontal_heading_cost	Return horizontal cost (1 - cos(theta))/2 and availability per action.
robust_temperature_logits	Return retained-state logits for an inspection-only temperature shadow.

1.1.1 horizontal_heading_cost

rollouts.inspection.selector_math.horizontal_heading_cost(
    candidate_centers_world,
    current_camera_center_world,
    previous_camera_center_world=None,
    displacement_tolerance_m=0.0001,
)

Return horizontal cost (1 - cos(theta))/2 and availability per action.

1.1.1.1 Parameters

Name	Type	Description	Default
candidate_centers_world	torch.Tensor	Tensor["N 3", floating] camera centers in world metres.	required
current_camera_center_world	torch.Tensor	Tensor["3", floating] last camera center.	required
previous_camera_center_world	torch.Tensor | None	Tensor["3", floating] preceding camera center, absent for first acquisition. Supply the calibrated root camera center, rather than the corrected reference rig center.	None
displacement_tolerance_m	float	Horizontal lengths at or below this metre threshold have unavailable heading, including zero/vertical moves.	0.0001

1.1.1.2 Returns

Name	Type	Description
	torch.Tensor	Tensor["N", floating] costs in [0, 1] and Tensor["N", bool]
	torch.Tensor	availability on the candidate device. Straight, perpendicular and
	tuple[torch.Tensor, torch.Tensor]	reversed displacements cost 0, 0.5 and 1. Missing history or nonfinite
	tuple[torch.Tensor, torch.Tensor]	centers yield neutral zero cost with false availability. World Z is up;
	tuple[torch.Tensor, torch.Tensor]	joint XY rotations/reflections preserve cost, independently of gaze.

1.1.2 robust_temperature_logits

rollouts.inspection.selector_math.robust_temperature_logits(
    scores,
    temperature,
    iqr_floor=0.0,
    heading_cost=None,
    heading_weight=0.0,
)

Return retained-state logits for an inspection-only temperature shadow.

The nonnegative iqr_floor bounds score amplification when an admitted candidate set has a nearly collapsed score distribution. A zero floor retains the historical dtype-epsilon stabilization exactly. Enabled heading shaping uses logits = (z - weight * cost) / temperature without rescaling combined utility. Scores and aligned costs are Tensor["N", floating] on the same device. Nonfinite rewards retain unavailable NaN logits. Positive weight requires supplied, finite costs in [0, 1] with the exact score shape/device. Temperature is finite and positive; weight and scale floor are finite and nonnegative. Controls must be representable in the score dtype; nonfinite results for finite scores raise ValueError. Normalization delegates to the canonical replay calculation through producer_robust_temperature_logits.