SCOPE FIRST
Different robot applications fall under different safety standards and risk contexts.
ISO 10218-1:2025 covers safety requirements for industrial robots themselves, while ISO 10218-2:2025 focuses on integration of industrial robot applications and cells. Collaborative industrial operation has additional guidance in ISO/TS 15066. Other robot classes can fall under different standards and regulatory regimes.
OpenRGD does not certify conformance to those standards and a valid RGD profile is not evidence of certification.
MULTILAYER SAFETY
Physical AI adds new reasoning layers without removing classical safety responsibilities.
Inherently safe design
Mechanical, electrical and machine-level risk reduction belongs to the robot/system design.
Application integration
Cells, tools, environments, commissioning and operating procedures create system-level hazards.
Perception + semantic safety
AI systems may detect humans, hazards, prohibited targets or ambiguous instructions.
Action validation
Proposed actions can be checked against capabilities, policies and declared operational constraints.
Runtime + control safety
Deterministic limits, watchdogs, state machines, fallback and emergency behavior belong near execution.
Evidence + audit
Structured traces can record which policy, evidence and state contributed to an execution decision.
AI SAFETY ≠ FUNCTIONAL SAFETY
Semantic understanding can complement, not replace, deterministic safety mechanisms.
Current robotics platforms increasingly add AI-based safety perception and reasoning, while model providers publish safety evaluations and limitations for embodied systems. Those mechanisms can improve situational awareness and instruction handling, but they do not erase the need for application-specific risk assessment, functional safety engineering and physical protective measures.
OPENRGD POSITION
OpenRGD can make pre-actuation constraints inspectable.
The canonical OpenRGD standard can represent operational context and constraints. Separately, the current non-normative agent contracts explore an Operation Safety Gate and DecisionTrace between a hardware-agnostic CapabilityPlan and the Body Adapter. This is an architectural boundary for evidence and action validation — not a safety certification.
FAIL CLOSED
Uncertainty should remain visible at the boundary.
OpenRGD’s broader design philosophy is to preserve missing or uncertain evidence rather than silently manufacture defaults. Applied to safety architecture, that favors explicit escalation, blocking or reduced authority when required body, environment or policy evidence is absent — while leaving real-time protective control to the systems designed and certified for it.