SILVER/NODEDC_MISSION_CORE
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

feat(perception): define semantic object understanding

Browse Source
This commit is contained in:
DCCONSTRUCTIONS
2026-08-25 16:43:47 +03:00
parent 1084984da2
commit 6276bbf324
28 changed files with 7028 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/domain-model/mission-core-experimental-vocabulary-v0alpha2.md
+42
View File
@@ -75,6 +75,48 @@ device protocol, transport, identity, or command state.
12. A plugin UI contribution is not ownership of Mission Core navigation,
scene layout, persistence, or global application state.
## Object understanding projection
Mission Core also has one bounded, executable perception vocabulary profile:
`missioncore.urban-object-semantics/v0`. It is referenced by the separate
`missioncore.object-understanding/v1` document and remains experimental. This
is a versioned projection over an existing `ObstacleObservation`; it is not a
second occupancy model, a long-term identity store, or a runtime ontology
service.
| Term | Meaning |
|---|---|
| `ObjectUnderstanding` | A source/frame-bound composition of immutable geometry, ranked semantic hypotheses, explicit semantic resolution, object state, advisory risk and evidence provenance. |
| `SemanticHypothesis` | One provider label mapped to a canonical class candidate with confidence, rank and evidence IDs. It is not class truth. |
| `SemanticDecision` | An explicit `unresolved`, `selected`, `ambiguous` or `conflict` result over the ranked hypotheses. |
| `ObjectStateEstimate` | Observed motion plus a separately qualified agency claim. A class prior is carried as a prior and never rewritten as observed motion. |
| `AdvisoryRiskAssessment` | Evidence-qualified risk level and suggested response for a future policy consumer. It never grants planner, navigation, safety or actuation authority. |
| `EvidenceProvenance` | The exact provider/model/revision/preprocess/prompt-set identity and source-frame binding supporting a claim. |
The v0 class profile is intentionally shallow. It covers the current named
consumers: human/adult/child, animal/dog, vehicle subtypes, passive wheeled
objects, common small static street obstacles, terrain discontinuities and
natural debris. Unknown and group-level classes remain valid outcomes; the
profile does not require a fragile leaf-class decision.
Mandatory object-perception non-conflation rules:
1. Occupancy identity is owned by `ObstacleObservation`, not by a semantic
class, detector label or provider tracklet.
2. A missing or unresolved semantic class never means free space and never
removes an occupied obstacle.
3. A raw provider label is not a canonical class; a ranked class hypothesis is
not a selected class; a selected class is not independent truth.
4. Observed motion is not a class motion prior. In particular, a currently
stationary vehicle may retain a self-propelled agency prior.
5. Semantic risk is advisory. `reduce-speed`, `yield`, `stop` and
`route-around` values describe policy hints, not issued commands.
6. Vehicle-dependent traversability is not encoded as an intrinsic property
of a curb, step, hemisphere or other obstacle class.
7. The class-free M4.8 obstacle-quality truth remains class-free. Semantic
evaluation requires a separate sealed label generation and cannot rewrite
an existing M4.8 result.
## Deferred compatibility work
The following meanings are intentionally unresolved in v0alpha2. They must be