1The Missing Abstraction

Here, Physical AI refers broadly to computational systems that perceive, reason about, or act upon physical environments. It is usually described in terms of perception, world models, planning, control, and manipulation. Those capabilities answer important questions: What is present? What is happening? What should be done? How should an actor move? Yet many real tasks contain another question that is logically distinct.

Which physical resource should satisfy this objective, where is its actionable endpoint now, and is that resolution still valid?

Today, that question is often answered implicitly. A robot may infer it from perception. A warehouse system may look it up in inventory. A technician may know it from memory. A digital twin may expose it through an application-specific interface. A planner may assume it has already been solved. Treating these mechanisms as equivalent obscures the common computational problem beneath them.

IAPI-06 introduced Intent-to-Physical-Resource Resolution (IPRR) as that problem class [1]. The purpose of this paper is to sharpen the claim: IPRR is not merely a useful function inside one architecture. It is a candidate infrastructure primitive for Physical AI because the binding between objective and physical endpoint must be performed repeatedly, under changing state, by heterogeneous actors and systems.

2Problem Definition

Let a request be q = (i, c, t, m, τ_req), where i denotes intent or objective; c denotes bounded context and policy; t is decision time; m is the anticipated response mode; and τ_req is the minimum validity horizon required by the requester. Let Ω(q) denote the set of acceptable physical solution tuples ⟨r, x, s, a⟩ for q, where acceptability may encode capability, policy, operational state, spatial adequacy, and endpoint actionability.

A resolver maps q either to a bounded resolution assertion ρ(q) = ⟨r_i, x_i, s_i, a_i, τ_asserted⟩ or to the explicit abstention outcome ABSTAIN. Here r_i is a selected physical resource instance; x_i is relevant spatial state; s_i is relevant operational state; a_i is an actionable physical endpoint; and τ_asserted is the interval for which the resolver is prepared to support the assertion.

Resolution is admitted only when ⟨r_i, x_i, s_i, a_i⟩ ∈ Ω(q) and τ_asserted ≥ τ_req. Otherwise ρ(q) = ABSTAIN.

Admission rule

This formulation makes a physical resolution fundamentally different from a static lookup. The result is not merely an identity or coordinate. It is an assertion that a particular resource-state-endpoint tuple is acceptable under the current context and that the assertion is expected to remain usable for at least the requested interval.

Correctness is spatiotemporal and state-dependent

A resource can be correctly identified yet incorrectly resolved. It may have moved; become occupied, depleted, unsafe, incompatible, unauthorized, or unreachable; or remain present while its actionable endpoint has changed. Physical correctness therefore depends on more than identity. It depends on the relationship among objective, resource, state, space, endpoint, and time.

The distinction between τ_req and τ_asserted is central. τ_req belongs to the consumer: how long must the answer remain usable? τ_asserted belongs to the resolver: how long can the answer be supported by current evidence? When the second is shorter than the first, abstention is preferable to false certainty.

3Search, Resolution, and the DNS Analogy

The phrase "the DNS of Physical AI" is intentionally compressive, but the analogy has a precise boundary. DNS does not infer what a user means. It accepts a name and resolves it to records useful for reaching a networked endpoint. IPRR may begin earlier, with an objective expressed above the level of a resource identifier or physical address.

objective → semantic requirement → eligible set → instance → state/space validation → endpoint → validity assessment → admission

IPRR resolution chain

The left side of this chain is search-like: it interprets an objective and narrows a candidate space. The right side is resolution-like: it binds an admissible physical target to an actionable endpoint. The complete problem therefore combines semantic discovery with endpoint resolution. The analogy becomes strongest after semantic requirements are sufficiently specified. At that point, the requester should not need to know the underlying physical endpoint in advance. This is the same architectural benefit that makes resolution layers powerful in digital systems: they decouple the consumer's objective from low-level addressing.

DNS resource records already include a time-to-live (TTL) value that specifies how long a record may be cached before it should be discarded [9]. That resemblance is useful but limited. DNS TTL governs reuse of a record; it is not conditioned on a requester's required horizon and does not assert that a mutable physical resource will remain admissible, reachable, or actionable throughout that interval. IPRR therefore separates the consumer's τ_req from the resolver's τ_asserted.

The analogy also breaks in an important way. Network names do not ordinarily become unavailable because another agent picked them up, moved them, depleted them, or made them unsafe. Physical endpoints are mutable and contestable. IPRR is therefore better understood as resolution under physical state, not as DNS transplanted into the physical world.

4What Existing Systems Resolve

The claim for IPRR is not that resolution around physical objects has never existed. It has. The relevant question is what each neighboring system takes as input, what it returns, and what correctness condition it enforces.

• EPCglobal Object Name Service. ONS is the most important precedent for any DNS analogy involving physical objects. The ratified EPCglobal ONS specification used DNS machinery to transform an Electronic Product Code into queries whose results point toward services associated with that EPC [3], and the EPC identifier family was registered as a URN namespace in RFC 5134 [8]. But ONS begins after identity is known. IPRR begins one abstraction earlier: given an objective, which physical resource instance is admissible? The difference is not old DNS versus new DNS; it is identifier resolution versus objective-to-physical-resource resolution.
• GS1 Digital Link and conformant resolvers. GS1 Digital Link provides standardized URI syntax for GS1 identification keys, while the GS1-Conformant Resolver standard resolves those identifiers to one or more related information sources or services [4][5]. The model is scalable and federated, but its input remains an identified entity. IPRR addresses the unresolved binding from objective to admissible physical entity and actionable endpoint.
• CoRE Resource Directory. The IETF CoRE Resource Directory addresses discovery where direct discovery is impractical: servers register resource information, and clients look up links describing resources [6]. This is highly relevant infrastructure for constrained environments, but it does not specify the correctness relation that makes a physical target admissible for a particular objective over a required validity horizon.
• Intentional Naming System. INS is conceptually closer because applications describe what they seek rather than where a service is located [2]. It was designed for dynamic and mobile networks and explicitly framed names around intent. IPRR extends the question into the physical domain, where the target is not merely a reachable service: its location, operational state, physical adequacy, contention, and actionability can change independently of identity.
• Multi-robot task allocation. MRTA formalizes a different assignment problem: which robot or robots should perform which task or tasks [7]. IPRR instead asks which physical resource can satisfy the objective and where its actionable endpoint is. The two can be composed, but neither subsumes the other.
• Semantic mapping, grounding, and affordance. Hydra constructs hierarchical 3D scene graphs containing objects, places, rooms, and their spatial relations [10]; ConceptGraphs supports open-vocabulary 3D scene representations queried by abstract language prompts for downstream planning [11]; and SayCan grounds language-model suggestions through robot skills and affordance-related value functions [12]. These overlap parts of the IPRR chain, especially semantic interpretation, physical grounding, and feasibility. The IPRR claim is not that objective-to-instance grounding is absent from robotics, but that the broader resolution contract, admissibility under current physical state, an actionable endpoint, explicit validity horizons, abstention, and potentially cross-domain authority, can be isolated as a common problem independent of the mechanism that performs the grounding.

5Why Physical Resolution Is a Distinct Problem Class

A problem class becomes useful when different implementations can be compared against common semantics. IPRR supplies such semantics by separating the resolution question from where computation occurs. A human, robot-local perception stack, centralized application, intelligent environment, distributed infrastructure service, or hybrid system can all attempt the same mapping q → ρ(q). This separation exposes at least six properties that are otherwise easy to hide inside application logic.

1. Representation and sensor heterogeneity. The same physical state may be represented through UWB, RFID, BLE, machine vision, fixed automation systems, digital twins, or combinations of these. IPRR should not require a resolver to be rewritten for every sensing substrate. The problem-class abstraction instead requires heterogeneous observations to support a common resolution semantics: resource identity and capability, spatial state, operational state, endpoint actionability, provenance, confidence, and freshness.
2. Admissibility. The nearest or most obvious resource is not necessarily acceptable. Compatibility, policy, authorization, condition, capacity, workflow, and task-specific constraints determine Ω(q).
3. Spatial validity. Location is not metadata appended after selection; it is part of whether a resolution can be acted upon. The required spatial representation may range from a facility or zone to a shelf position, pose, approach vector, or other endpoint geometry.
4. Temporal validity. Physical state decays. A correct answer at t may be wrong at t + Δ. The validity horizon converts freshness from an implementation detail into part of the resolution contract. An admitted tuple remains valid only while it continues to satisfy the request constraints represented by Ω(q) throughout [t, t + τ_req].
5. Contention. Physical resources can be exclusive or consumable. A resolution may become invalid because another actor claims, moves, depletes, or blocks the selected target. Coordination mechanisms may exist outside the resolver, but their effects must be reflected in whether the resolver can sustain τ_asserted.
6. Abstention. A system that cannot support the requested horizon should be allowed to return ABSTAIN. This is not failure in the ordinary sense; it is a correctness-preserving outcome. In safety- or mission-relevant environments, explicit abstention is preferable to returning a physically stale or weakly supported endpoint.

6Federation, Authority, and Trust

If IPRR remains local, each deployment can define its own semantics. If it becomes infrastructure, independently governed physical domains must eventually interoperate. The key scaling question is not whether the world should maintain one global map of every object. It is whether authoritative domains can answer bounded resolution questions about resources they govern.

A federated model suggests several requirements: domain discovery; semantic interoperability; delegation of authority; authentication of the asserting domain; authorization of the requester; integrity and provenance of state assertions; freshness and replay resistance; and safe abstention when trust or evidence is insufficient. This paper does not propose a new federation or security protocol. The point is architectural: once a resolution can cross administrative boundaries, trust becomes part of the resolution problem. A physically plausible endpoint from an unauthenticated or unauthorized authority is not an acceptable resolution.

Authority transfer and mobility

Physical resources may cross sensing, operational, or administrative boundaries during the lifetime of a resolution. Federation must therefore distinguish persistent resource identity from the domain currently authoritative for the resource's spatial and operational state. A resource can remain the same instance while the authority capable of making a trustworthy assertion about that instance changes.

If authority transfers from domain D_A to domain D_B, a previously admitted resolution may require expiration, revalidation, or an explicit transfer of authoritative state. A resolver should not assert a validity horizon that extends beyond the interval for which it has authoritative knowledge of the selected resource. In this sense, mobility is not merely a tracking or handoff problem; it is a resolution-validity problem. The requirement is architecture-neutral. A future implementation might use hierarchical directories, delegated authorities, distributed registries, peer-to-peer exchange, or mechanisms not yet defined. IPRR specifies the invariant: the authority supporting a resolution must remain sufficient for the duration of the asserted horizon, or the resolution must be revalidated or withdrawn.

7Evaluation: What Would It Mean to Solve IPRR Well?

Treating IPRR as a problem class allows implementations to be evaluated without prescribing their architecture. Candidate metrics include:

• Resolution correctness. Fraction of admitted resolutions contained in Ω(q).
• Coverage. Fraction of requests for which the system admits a resolution rather than abstaining.
• Horizon reliability. Fraction of admitted resolutions that remain valid through τ_req.
• Horizon calibration. Agreement between τ_asserted and observed resolution survival, measured across predicted horizon bands so that asserted validity is neither systematically overconfident nor unnecessarily conservative.
• Resolution latency. Time from q to admitted ρ(q) or ABSTAIN.
• Spatial adequacy. Whether the returned endpoint is precise enough for the intended response mode m.
• State freshness. Age and confidence of the evidence supporting x_i and s_i.
• Contention robustness. Degradation in correctness or coverage as simultaneous demand increases.
• Interoperability. Ability to resolve across heterogeneous representations, sensing systems, and administrative domains.

Operationalizing Ω(q) requires ground truth. In controlled evaluations, admissibility can be labeled from known policy and state; in deployed settings, it can be assessed through endpoint verification, event logs, or post-hoc audit of whether the selected resource-state-endpoint tuple actually satisfied the request constraints at decision and execution time. These metrics expose tradeoffs. A resolver can increase coverage by admitting weaker assertions, or increase reliability by abstaining more often. The correct operating point depends on the domain, but the tradeoff becomes measurable once the resolution problem is isolated.

8Research Agenda

Several research questions follow directly from the problem definition. How should physical capabilities, eligibility constraints, spatial state, and actionable endpoints be represented across heterogeneous domains? How should τ_asserted be estimated and communicated? What evidence is sufficient to support an assertion over a requested horizon? How should contention alter admissibility? When should resolution remain actor-local, when should it move into the environment, and when should it be federated? How should semantic registries evolve without forcing every actor to share an identical ontology? How should authority transfer when resources cross physical or administrative boundaries, and how should outstanding resolution assertions be invalidated, transferred, or revalidated during that handoff?

A second set of questions concerns composition. IPRR can sit between high-level planning and downstream execution, but it can also interact recursively with task allocation, motion planning, inventory systems, and digital twins. The important boundary is conceptual: IPRR ends when an actionable physical endpoint has been admitted. What happens at that endpoint belongs to execution.

9Conclusion

Physical AI has largely treated resource resolution as an implementation detail distributed across perception, inventory systems, planning, localization, and human knowledge. We argue that this is a category error. Resolving an objective into an admissible, spatially valid, temporally valid physical endpoint is itself a computational problem. Once isolated, it can be specified, measured, compared, and implemented independently. We call that problem Intent-to-Physical-Resource Resolution.

The Internet required resolution infrastructure because applications could not scale by knowing every network endpoint in advance. Physical AI will confront an analogous constraint in the physical world. Its endpoints, however, move, change state, become contested, and expire. The resulting problem is not DNS for objects. It is resolution under physical state.

Before the physical world can become programmable, it must first become resolvable.