SECTION 01
A VLA output can be useful without being authoritative.
Different VLA systems expose different action interfaces. Some produce high-level skills, some trajectories, some discretized actions and some low-level control-like outputs. OpenRGD does not impose one universal VLA architecture. The candidate boundary is instead about authority: whatever the model emits should enter the embodied system as a proposal whose meaning is validated against the actual body.
That distinction protects portability. A model can remain model-centric while embodiment-specific constraints remain attached to the robot, runtime and application.
SECTION 02
The candidate path inserts explicit transformation and validation.
In the current OpenRGD architecture work, CognitionProposal becomes ActionIntent before somatic translation. The Somatic Translator resolves the intent against body semantics and produces a CapabilityPlan. An Operation Safety Gate then evaluates whether that plan can proceed under the current constraints before the Body Adapter reaches hardware.
DecisionTrace sits in the path as an auditable record of the decision context. This does not make the architecture a certified safety system; it makes the semantic transitions inspectable.
CognitionProposal → ActionIntent → Somatic Translator → CapabilityPlan → Operation Safety Gate → DecisionTrace → Body Adapter → HardwareSECTION 03
ActionIntent is a semantic boundary, not a new control protocol.
ActionIntent gives the model-facing side a stable way to express what is being attempted without requiring the model to own device IDs, bus protocols, calibration tables or controller-specific details. It describes intended embodied action at a level that downstream systems can validate and translate.
That also gives multiple cognition providers a common handoff point. A VLA, planner, policy or world-model-backed agent can differ internally while still producing a contract that the embodied runtime understands.
SECTION 04
Somatic translation is where generic intent meets a specific body.
The same request can imply very different feasible actions on different robots. Somatic translation uses the body graph and capability semantics to turn a generic intent into a plan compatible with the current embodiment. Unsupported resources, missing prerequisites or ambiguous mappings should surface here instead of being hidden in a controller.
This is one reason OpenRGD invests in capability semantics and provenance rather than only structure. Translation is only as trustworthy as the body context it consumes.
SECTION 05
The safety gate must remain separate from the model's confidence.
Model confidence, language plausibility and policy likelihood are not equivalent to operational safety. A runtime safety layer needs its own constraints, timing, state and failure behavior. In a physical system, some of those mechanisms may need deterministic execution far outside the latency envelope of a large model.
OpenRGD therefore treats semantic safety as one layer in a larger robot safety architecture. The candidate contracts can help explain what was requested and why a plan was considered, but functional safety remains a system-level responsibility.
SECTION 06
Direct-control VLAs do not invalidate the boundary.
Some research systems intentionally map observations directly to low-level actions. OpenRGD does not claim that such systems are conceptually impossible or universally wrong. The architectural question is whether a production embodied system wants the model to own final actuator authority.
Where direct control is appropriate, the surrounding runtime still benefits from explicit body identity, execution context, hard constraints, monitoring and traceability. The exact contract can vary; the need to make authority visible does not.
SOURCES & PRIMARY CONTEXT