Namespaced ids prevent Galbot G1, Unitree G1, and other collisions from sharing claims.
Evidence before edges
Every edge carries status, scope, boundary notes, review dates, claims, and source records.
Data-release events are immutable. Relation observations preserve validity and review dates.
Coverage, evidence, access, deployment, and interoperability remain separate dimensions.
Relationship explorer
Select an exact entity and filter its reviewed inbound and outbound relations. The default is contextual adjacency—not a force graph—so labels, evidence, and negative boundaries remain readable.
Physical AI stack workspace
Trace up to two reviewed hops from a robot through organizations, models, data, runtimes, simulators, compute, and operations. Boundary edges are isolated from the stack so “not proven on” can never be rendered as compatibility.
Temporal signals, not a trend score
The observatory reports literal dimensions and immutable data-release activity. These counts describe what is evidenced in this dataset; they do not estimate market share, technical quality, safety, or likely winners.
Data release ledger
Recently observed relations
Deployment and openness ledger
Categorical states stay independent. Public weights do not establish production support; a real-robot paper evaluation does not establish commercial deployment; unknown remains unknown.
Loading the evidence graph…
Agent-native endpoints
Stable nodes, scoped relations, dimensions, release events, and compact source records.
One static HTML page and one JSON document per graph entity, including canonical claims and adjacent relations.
Line-delimited release and relation-observation records for incremental ingestion.
All source-table claims remain in the complete archive. The graph is a reviewed identity-and-relation projection.
Graph boundary
An absent edge means this review has not established the relation. It does not prove incompatibility. A claimed edge remains attributed; a boundary edge records a tempting but unsupported join. Entity dossiers use exact ids and authored bindings—never fuzzy names.