Bounded authority
Every autonomous action should be limited by explicit task, location, machine, time, environmental, and operating constraints.
Software can retry cheaply. Heavy equipment cannot. Armature's autonomy thesis starts from the reality that every command ultimately moves mass through a physical environment shared with structures, utilities, machines, and people.
The machine does not receive permission to “finish the job somehow.” It receives bounded authority tied to space, state, task, machine health, observed conditions, and a defined safe response when those assumptions stop being true.
Authority is explicit space plus explicit conditions. When perception, localization, communication, or human separation leaves that envelope, the valid action set should shrink toward a safe state.
The objective is not an AI that can always invent a next move. It is a system that knows when it is authorized to act, when it must stop, and what evidence is required before downstream work proceeds.
Every autonomous action should be limited by explicit task, location, machine, time, environmental, and operating constraints.
The system should know what it has actually observed, what it inferred, how current that state is, and where uncertainty remains.
Loss of confidence, communications, required sensing, authorization, or safe operating conditions should drive the machine toward a defined safe state.
Critical completion criteria should use evidence capable of detecting a task that stopped without actually succeeding.
Humans retain explicit paths to pause, isolate, override, inspect, and change objectives; autonomy does not erase operational responsibility.
New capability should expand from constrained environments only after its assumptions, failure modes, and recovery paths are understood.
A SAFE SYSTEM MUST BE ABLE TO DISTINGUISH SUCCESS FROM “DID NOT RUN.”
Verification is not a green process indicator. Physical outcomes need positive controls: measurements and observations that would differ if the intended work had not actually occurred.
Each layer should narrow intent into a smaller, inspectable authority envelope instead of passing an unconstrained objective directly to machinery.
Human-approved engineering goal and acceptance criteria.
Sequenced work with dependencies, boundaries, and assumptions.
Specific machine, area, tool, time, and completion conditions.
Locally feasible trajectory inside current observed conditions.
Low-level control constrained by machine safety systems.
Independent observation that closes or rejects the task.
A robust autonomy system assumes sensors fail, maps become stale, terrain changes, networks disappear, people enter the work zone, and physical conditions diverge from expectation.
Observed geometry or objects disagree with the expected world model. Action should pause or narrow until the discrepancy is resolved.
Machine position or orientation is not known precisely enough for the task. The acceptable uncertainty depends on the physical consequence of error.
A machine cannot rely on a remote coordinator being continuously reachable. Local safe-state behavior must remain defined when the network disappears.
Braking, steering, hydraulics, power, attachments, sensors, and structural state can deteriorate. Plans must respect actual machine health, not nominal capability.
People, emergency responders, contractors, and the public can violate planned separation. Detection and immediate protective behavior take priority over productivity.
If evidence cannot confirm completion, the system must not silently promote the task to success just because execution returned without an error.
Subsurface autonomy can couple air, water, access, fire, and localization failures into the same event. Consequence must expand more slowly than capability: each hazard needs observable state, a bounded machine response, and a human authority boundary before autonomous work is permitted.
Failure: airflow, pressure control, smoke extraction, or a required ventilation path becomes unavailable or unreliable. Observe: airflow and pressure state, smoke, atmosphere, equipment state, and sensor freshness. Autonomous authority: stop affected work and enter only a predefined local safe state; isolation is allowed only where its effects are already bounded. Human authority: decide evacuation, ventilation recovery, and re-entry. Proving safe airflow across connected underground volumes remains an engineering validation problem.
Failure: heat, smoke, or fire threatens a volume with limited exit routes. Observe: fire indicators, smoke conditions, occupied areas, route availability, and the state of local power and barriers. Autonomous authority: stop work, avoid routing machines into the affected volume, and execute only pre-authorized protective actions. Human authority: evacuation, rescue, suppression strategy, and re-entry remain under people and emergency responders; autonomous fire response is not claimed here.
Failure: water enters faster or somewhere other than the model expects, or drainage and pumping capacity becomes uncertain. Observe: water level and rate of change, ingress location, pump and valve state, power availability, and downstream capacity. Autonomous authority: stop excavation in the affected zone, block new machine entry, and retreat only where a validated route remains available. Human authority: dewatering, structural assessment, changed ground assumptions, and re-entry require explicit review.
Failure: excavation response, lining behavior, settlement, deformation, or adjacent ground movement leaves the validated model. Observe: task-appropriate geotechnical and structural measurements, excavation state, lining state, displacement trends, and evidence freshness. Autonomous authority: stop cutting, prevent entry into the affected zone, and withdraw only where the route remains demonstrably safe. Human authority: stabilization, changed support assumptions, and permission to resume require site-specific engineering review. General autonomous recovery from unstable ground is an open engineering problem.
Failure: a person or machine is trapped, injured, blocked, or cut off from a usable access route. Observe: known occupancy, last reliable location, route clearance, machine positions, and structural or environmental conditions along the access path. Autonomous authority: stop, yield, and clear only predefined rescue access where doing so cannot increase consequence. Human authority: rescue command and movement around a casualty remain human-controlled. Autonomous rescue is not a solved capability.
Failure: oxygen, toxic or flammable gas, dust, or another atmospheric condition leaves the authorized envelope. Observe: task-appropriate atmospheric sensing, ventilation state, spatial coverage, calibration status, and reading freshness. Autonomous authority: halt affected work and move to a predefined safe condition only while the required sensing and route remain valid. Human authority: classify the hazard, set protective controls, authorize purge or inspection, and permit re-entry.
Failure: one module bay cannot be made safe without transferring its hazard or dependency failure into the corridor network. Observe: physical boundaries, utilities, airflow, traffic, material routes, and every dependency crossing the isolation boundary. Autonomous authority: when a bounded trigger requires the bay to be made safe, a machine may execute a predefined isolation sequence, positively verify the resulting state, and prevent new routing into the module bay. Human authority: a person must authorize restoration, re-entry, or any changed network topology after reviewing the isolation state. Proving that isolation does not move the failure elsewhere is an open engineering problem.
Failure: communications degrade or the machine can no longer localize accurately enough for its task in an environment without dependable GPS or line of sight. Observe: network health, local position uncertainty, map agreement, reference availability, and state age. Autonomous authority: shrink the action envelope as uncertainty grows and stop before it exceeds the task limit; return movement is permitted only when localization and the route remain valid. Human authority: recovery and resumed work require explicit re-authorization after trustworthy state is restored.
The machine performing work and the system deciding that the work is acceptable should not share every failure mode.
Use survey, lidar, vision, force, weight, position, or other task-appropriate evidence to compare actual output with acceptance criteria.
Completion should preserve enough state to reconstruct what changed, which machine acted, what model version was used, and what evidence closed the task.
Conflicting sensors or validation results should remain visible as unresolved state rather than being averaged into a false appearance of certainty.
Validation needs positive controls demonstrating that the check rejects a deliberately wrong or absent result instead of merely executing successfully.
Armature's development path should widen autonomy from controlled tasks to larger physical systems only when the previous scope has measurable operating boundaries and recovery behavior.
Talk with Armature →