TERMINOLOGY
Capability, skill, task, action and affordance should not be treated as interchangeable.
Capability
A declared ability of the embodiment under specified prerequisites and constraints.
Skill
A reusable behavior or policy that may realize one or more capabilities.
Task
A goal-oriented unit of work to be achieved in a context.
Action
An explicit operation or semantic intent considered for execution.
Affordance
A relation between agent/body and environment that makes an action possible.
Resource
Sensors, actuators, body parts, software or context required to realize the capability.
GROUNDING
A useful capability should be testable against the actual embodiment.
KnowRob/SRDL has long represented robot hardware, software and capabilities and can check whether dependencies of action descriptions are available on a robot. OpenRGD’s current Agency material explores a similar practical problem from an embodiment-profile perspective.
OPENRGD CURRENT WORK
The repository already exposes useful capability primitives — with mixed maturity.
The current skill schema includes required sensors and actuators, world-model preconditions, safety envelopes, control profiles, runtime learning policy and fallback behavior. Separately, the candidate Somatic Translation contract defines a hardware-agnostic CapabilityPlan.
These artifacts are valuable design evidence, but they do not all share the same maturity. The CapabilityPlan contract remains candidate, and OpenRGD.org should present capability modeling as an evolving architectural area rather than a finalized universal ontology.
WHY VLA NEEDS THIS
Generalist models still need embodiment-specific grounding.
A model can propose “pick up the box,” but a heterogeneous robot fleet needs a separate answer to: which embodiment has the sensors, reach, gripper, payload, safety envelope and context required to perform that action? Capability models make that compatibility problem explicit.