1 targets

targets

Sanitized target instructions shared across ARIA-NBV owners.

1.1 Package guide

aria_nbv.targets owns actor-safe target instructions. It converts observed detector OBBs into stable descriptors without carrying ground-truth rows, matching diagnostics, labels, gains, or rollout storage details.

1.2 Public API

from aria_nbv.targets import observed_target_descriptors

observed = observed_target_descriptors(vin_offline_sample)
valid = tuple(item.descriptor for item in observed if item.descriptor is not None)

TargetDescriptor contains semantic identity, object pose, full metric extents, and the object pose relative to the snippet reference rig. ObservedTargetDescriptor additionally retains actor-visible source identity, confidence, and a provenance hash.

1.3 Ownership Boundary

  • targets.selection reads detected OBBs only and preserves their source-row ordering.
  • targets.protocol defines the target-input protocol and label-evidence admission vocabulary used by stores and QH readers.
  • oracle.target_selection performs privileged GT matching and task sampling.
  • Oracle labels, crop geometry, invalidity, and rollout persistence never enter TargetDescriptor.

The detailed field and frame contracts live in the public docstrings and generated API reference.

1.4 Verification

cd aria_nbv
uv run ruff check aria_nbv/targets
uv run pytest tests/targets tests/oracle/test_target_selection.py

1.5 Classes

Name Description
ObservedTargetDescriptor One actor-visible detected OBB and its stable descriptor identity.
TargetDescriptor Sanitized target instruction for target-conditioned candidate generation.

1.5.1 ObservedTargetDescriptor

targets.ObservedTargetDescriptor(
    sample_key,
    source,
    source_row,
    target_id,
    descriptor,
    confidence,
    inst_id,
    obb_data=None,
)

One actor-visible detected OBB and its stable descriptor identity.

No ground-truth row, match score, or privileged label is present. The descriptor_hash is intended for manifest/evidence provenance only.

1.5.1.1 Attributes

Name Description
sample_key Immutable source sample identity.
source Actor-visible OBB source label.
source_row Row index in the detected OBB table before padding removal.
target_id Stable identity for this detected source row.
descriptor Sanitized actor instruction, or None for retained invalid geometry.
confidence Detector confidence retained for audit.
inst_id Detector instance identifier.
obb_data Flattened actor OBB payload retained for privileged evaluation only.
descriptor_hash Return the canonical hash of actor-visible identity fields.

1.5.2 TargetDescriptor

targets.TargetDescriptor(
    sem_id,
    class_name,
    pose_world_object,
    extents_m,
    relative_pose_reference_object,
)

Sanitized target instruction for target-conditioned candidate generation.

The descriptor is a semantic task instruction. It may be GT-derived when an oracle specifies the task, but it does not carry GT source rows, matching diagnostics, crop state, labels, gains, invalidity, headroom, or persisted Zarr details.

1.5.2.1 Attributes

Name Description
sem_id Semantic class identifier for the target.
class_name Human-readable semantic class name.
pose_world_object Flattened 12-value object-to-world pose.
extents_m Full object side lengths in metres.
relative_pose_reference_object Flattened 12-value object pose in the snippet reference frame.
center_world Target center in world coordinates, metres.
target_id Return a stable actor-safe diagnostic id derived from descriptor fields.

1.5.2.2 Methods

Name Description
center_world_tensor Return center_world as a 3-vector tensor.
1.5.2.2.1 center_world_tensor
targets.TargetDescriptor.center_world_tensor(dtype=torch.float32)

Return center_world as a 3-vector tensor.

1.6 Functions

Name Description
observed_target_descriptors Convert detected OBB rows into deterministic actor-visible descriptors.

1.6.1 observed_target_descriptors

targets.observed_target_descriptors(sample)

Convert detected OBB rows into deterministic actor-visible descriptors.

Padded rows are omitted and source-row order is preserved. The function reads only actor-visible geometry, the neutral reference-frame pose, and sample identity; it never accesses the oracle payload or GT OBBs.

Back to top