Capability Awareness for Construction Robotics

Regulatory Framework

Construction robotics sits at the intersection of several distinct regulatory regimes, none of which were designed with mobile autonomous machines in mind. OSHA 29 CFR 1926 is the foundational federal standard for construction work in the United States, governing fall protection, struck-by hazards, scaffolding, excavation, electrical safety, and powered industrial trucks. Its provisions presume a human operator who exercises judgment about site conditions and assumes legal responsibility for compliance. When a robot performs the work, the chain of accountability is unclear: the contractor remains the cited party, but the operative judgments are made by software whose internal state is not visible to the OSHA-authorized competent person required by the standard.

ANSI/SAIA A92 governs aerial work platforms and mobile elevating work platforms, including the boom lifts and scissor lifts that are increasingly being adapted into autonomous or supervised-autonomous platforms. A92.20 imposes design requirements; A92.22 imposes safe-use requirements; A92.24 imposes training requirements. The training standard assumes a human occupant; when the platform carries a robotic end-effector instead, the standard provides no clear envelope. ISO 18497, originally developed for agricultural mobile machines, has been adapted as a reference for autonomous mobile equipment in unstructured outdoor environments, and its safety-related performance requirements increasingly inform construction practice. RIA/ANSI R15.06, derived from ISO 10218, governs industrial robots and integrators; its application to mobile construction robots is contested because R15.06 presumes a defined cell that does not exist on a job site.

IEC 62443 governs the cybersecurity of industrial automation and control systems and is the operative standard when construction robots integrate with site networks, building information modeling (BIM) servers, or general-contractor scheduling platforms. FAA Part 107 governs small unmanned aircraft used for site inspection, progress monitoring, and survey, and imposes operational limitations that interact with the construction schedule. The NIOSH Construction Safety Research agenda layers public-health surveillance and human-factors research onto the regulatory baseline. A capability-aware robot must produce evidence that satisfies all of these frameworks simultaneously, in real time, under conditions where the relevant rule set itself can change as the site moves between phases.

Architectural Requirement

Satisfying this fragmented regulatory regime is not a matter of bolting compliance checks onto a conventional autonomy stack. The architecture itself must treat the robot's capability as a first-class state variable that is continuously assessed, continuously bounded, and continuously communicated. The capability envelope is the formal expression of what the robot can presently do safely: which payload masses it can lift, which surface gradients it can traverse, which positional tolerances it can hold, which proximity it can maintain to human workers, and which environmental conditions degrade each of those parameters.

The envelope must be computed from sensor data describing the robot's actual mechanical, electrical, and computational state, fused with sensor data describing the environment, and projected forward across the time horizon of any task being considered. It must produce machine-readable outputs that downstream supervisory systems and human site managers can consume, and it must produce human-readable outputs in the form, vocabulary, and granularity that an OSHA competent person, an A92 platform supervisor, or a contractor's safety officer can act upon. The envelope is not a pre-computed table. It is a live computation whose value at any instant depends on the robot, the site, the weather, the workforce, and the task.

Architecturally, this requires three things that conventional construction robots lack. First, an introspection layer that observes mechanical wear, sensor degradation, computational backlog, communications latency, and power state. Second, a site model that is updated faster than the site itself changes, ingesting BIM updates, drone surveys, fixed sensors, and the robot's own perception. Third, a binding layer that combines the introspection state with the site state to produce a per-task capability assessment that can be stamped, logged, and presented as the basis for accepting, modifying, or refusing a work order. Without all three, capability awareness reduces to marketing.

Why Procedural Compliance Fails

Conventional approaches to robotic safety on construction sites rely on procedural compliance: a written method statement identifies the hazards, a risk assessment assigns mitigations, the robot is configured with conservative pre-set limits, and a human supervisor signs off. This approach is the natural extension of the OSHA competent-person model, and it works adequately when the site is static and the robot is performing a narrowly defined repetitive task. It fails when the site is dynamic, the task is varied, or the robot's capability changes within a shift.

The first failure mode is staleness. A method statement written on Monday assumes the floor configuration that existed on Monday. By Wednesday, formwork has been struck, MEP rough-in is partly complete, and the floor has new penetrations. The robot configured to Monday's plan is operating against a site that no longer matches its model. Procedural compliance has no mechanism for detecting this divergence; it relies on the supervisor to notice. On a multi-trade site with parallel work, the supervisor often cannot notice in time.

The second failure mode is conservatism collapse. Because pre-set limits must cover the worst case, they are routinely exceeded by site personnel who observe that the robot is operating well within its safe range and conclude that the limits are over-cautious. Override procedures accumulate. Within a few weeks, the procedural envelope on paper bears little resemblance to the operational envelope in practice. When a real edge case arises, the eroded margins do not protect the workforce.

The third failure mode is opacity. When an incident occurs, the post-event investigation must reconstruct what the robot believed about itself and the site at the moment of the incident. Procedural-compliance systems produce sparse logs: the work order, the configuration, the alarms, the disconnect. They do not produce a continuous record of capability state. OSHA citation defense, insurance subrogation, and the contractor's own root-cause analysis are all degraded by this opacity. The fourth failure mode is brittleness across regulatory regimes: a procedural envelope tuned to OSHA 1926 Subpart L scaffolding work does not automatically map to A92 aerial-platform work or to R15.06 enclosed-cell work, and the manual re-tuning at each transition introduces errors.

What AQ Primitive Provides

Adaptive Query's capability-awareness primitive provides the missing introspection-binding-projection stack as a single, audited construct. The primitive maintains a continuous capability state for the robot expressed as a structured envelope across the dimensions that matter to construction work: payload, reach, positional accuracy, force application, traversal, sensing, communication, and energy reserve. Each dimension carries a current value, a confidence bound, a degradation rate, and a forecasted value at the end of the next task and the end of the shift.

The primitive ingests robot-internal telemetry covering actuator current, joint backlash estimates, IMU drift, camera focus and exposure quality, network round-trip time, and battery state of health. It ingests environmental telemetry covering surface friction, slope, ambient light, precipitation, wind, dust loading, and detected human presence. It binds these inputs through a documented model that produces the envelope and a documented confidence in the envelope. The binding is deterministic, reviewable, and re-runnable against logged inputs, which is the property that distinguishes auditable capability awareness from opaque machine learning.

On task assignment, the primitive evaluates the task's requirements against the current envelope and returns one of three responses: accept with stated margins, accept conditionally with stated modifications, or refuse with stated cause. A robot whose positional accuracy has degraded due to vibration exposure during transport may lack the precision for structural placement but retain adequate capability for material delivery; the primitive expresses this as a refusal of the placement task with a specific cited dimension and a simultaneous acceptance of the delivery task. The site management system reassigns work accordingly. The robot continues contributing within its envelope rather than being parked.

The primitive also produces a continuous safety-margin computation for shared human-robot operation. When sensor capability is reduced by dust, rain, or sun glare, the safety perimeter expands and traversal speed decreases. When braking capability has degraded from operating on loose material, approach distances to human-occupied areas increase. These are not hand-tuned thresholds; they are derived from the same envelope used for task acceptance, which means a single coherent capability state governs both work and safety. The margins are always appropriate to actual capability, never the conservative fixed parameters that may be insufficient when capability has degraded or unnecessarily restrictive when capability is full.

Compliance Mapping

The capability envelope and its evidence stream map directly onto the obligations imposed by each regulatory regime that governs the work. Against OSHA 29 CFR 1926 Subpart C general safety provisions, the envelope provides a continuously updated competent-person-equivalent assessment, time-stamped and tied to the specific hazards identified in the contractor's site-specific safety plan. Against Subpart L scaffolding and Subpart M fall protection, the envelope binds the robot's traversal authority to verified surface and edge conditions. Against Subpart O motor vehicles and mechanized equipment, it binds traversal authority to verified ground bearing capacity and grade.

Against ANSI A92.22 safe-use requirements for aerial platforms, the envelope produces the equivalent of an occupant-state report for the robotic payload, including verified stability, wind exposure, and load distribution. Against ISO 18497, it produces the safety-related performance evidence for the autonomous-machine functions. Against RIA R15.06 and ISO 10218, it produces the speed-and-separation-monitoring data required for collaborative operation outside a fixed cell, and supplies the risk-reduction evidence that the standard requires the integrator to produce. Against IEC 62443, the envelope's communications and authentication state is itself part of the published capability, so a network-degraded robot voluntarily reduces authority rather than continuing to act on potentially compromised commands.

Against FAA Part 107 for inspection drones operating in the same airspace, the envelope coordinates with the drone's flight envelope to maintain separation and to suspend ground-robot operation during overhead inspection passes. Against the NIOSH Construction Safety Research agenda's emphasis on near-miss surveillance, the envelope's continuous record of margin consumption produces near-miss data automatically, without depending on worker self-report. Each regulatory citation has a corresponding envelope dimension, log field, and evidence retention path. The mapping is explicit, reviewable, and stable across the project's regulatory transitions.

Adoption Pathway

Adoption proceeds in three phases that match the risk tolerance of construction prime contractors and the maturity curve of construction robotics fleets. The first phase is shadow-mode instrumentation: the capability envelope runs alongside the existing autonomy stack, producing logs and dashboards but not gating any decision. The contractor's safety officer compares envelope output to observed events, tunes the binding model, and develops familiarity with the evidence stream. Shadow mode typically runs for a quarter on a representative sample of robots and tasks.

The second phase is advisory authority: the envelope's task-acceptance and safety-margin outputs become inputs to the supervisory system and to on-site personnel, but human override remains routine. During this phase the contractor refines the operating procedures around capability refusals, trains site supervisors to read the envelope outputs, and aligns the evidence retention schedule with the contractor's records-management policy. The owner's representative and the insurer are typically brought into the program at this stage.

The third phase is governing authority: the envelope gates task acceptance and continuously sets safety margins, with override available only through a documented procedure that itself is logged into the envelope's evidence stream. At this point the capability primitive is part of the contractor's compliance program, cited in the site-specific safety plan, and surfaced to OSHA inspectors and to the owner during commissioning. For construction companies facing severe labor shortages and an increasing share of work performed by mobile autonomous machines, this phased pathway provides the adaptive autonomous capability that dynamic construction environments demand without forcing a single high-risk transition. The robot does not need a perfectly modeled, perfectly controlled environment to operate productively under regulatory scrutiny. It needs to know its own capability and to assess whether that capability matches the conditions and tasks it encounters. Capability awareness provides precisely this structural self-knowledge, in a form that the prevailing regulatory frameworks can recognize and accept.