Task-scoped, not machine-wide
Autonomy receives explicit authority for a defined machine, tool, geometry, time window, dependencies, acceptance criteria, and stop conditions. It does not inherit unlimited control of the site.
Pilot 01 defines the smallest Armature field system worth proving: one machine, one useful earthworks task, one bounded operating cell, explicit human authority, and an independent evidence path that can reject the machine's own claim of success.
The pilot is intentionally constrained so failure cannot hide inside fleet complexity. Authority, geometry, sensing, intervention, material hand-off, and acceptance are visible in one physical loop.
One bounded physical task becomes a complete system only when authority, intervention, hand-off, measurement, and acceptance all survive the loop.
A ROBOT DEMO SHOWS THAT MOTION IS POSSIBLE. A FIELD CELL MUST PROVE THAT THE RESULT CAN BE TRUSTED.
The important milestone is not an excavator moving without a driver. It is a bounded system that can receive intent, refuse invalid authority, execute useful work, stop safely, measure reality, and preserve evidence.
Each phase adds a new class of risk or coordination. The system should earn more authority through evidence instead of treating autonomy as a single on/off capability.
Run the complete task contract against recorded and synthetic site states. Prove planner refusal, stop conditions, stale-state rejection, recovery paths, and evidence generation before physical authority exists.
Operate the machine conventionally while Armature observes site state, machine state, task progress, survey references, material movement, and acceptance measurements. Establish the physical truth the autonomous system must match.
Allow bounded machine execution with a human operator controlling task release and intervention. The system must halt safely on uncertainty, boundary conflict, perception disagreement, or loss of required state.
Measure grade, volume, geometry, and task completion using an evidence path independent of the machine's own completion signal. A stopped or non-executed task must be distinguishable from success.
Repeat the same class of useful task across changing site conditions without rebuilding the operating model from scratch. Compare productivity, intervention rate, error, and evidence quality over time.
Introduce one downstream machine or logistics hand-off only after the single-machine cell can preserve task state, authority, material identity, and acceptance criteria reliably.
The pilot treats authority as a versioned object with physical limits. A valid plan is not enough; the system must still know where, when, with what state, and under whose release it is allowed to act.
Autonomy receives explicit authority for a defined machine, tool, geometry, time window, dependencies, acceptance criteria, and stop conditions. It does not inherit unlimited control of the site.
The work zone, transit path, human access points, exclusion areas, utilities, and no-authority regions must exist in the shared site model before motion is released.
Plans are bound to an identified site-state revision. Material changes, survey updates, machine state changes, or conflicting observations can invalidate that authority and force replanning.
Emergency stop, supervised pause, remote/manual recovery, and safe re-entry are designed parts of the operating model—not exceptional behavior added after autonomy works.
The evidence plane exists so downstream work never has to infer success from a command having been sent or a machine reporting that it finished.
The task contract defines the measurable physical outcome before execution starts: target grade, trench geometry, volume, tolerance, placement, or another externally observable result.
Machine state, plan revision, authority, interventions, stops, retries, and task transitions remain attributable to one execution attempt.
Acceptance data comes from a measurement path that can disagree with the machine or planner. Agreement is evidence; automatic self-confirmation is not.
The system must preserve enough information to explain non-execution, partial completion, boundary refusal, uncertainty, or verification failure without rewriting history as success.
Pilot 01 should graduate because the system is measurably trustworthy at its current boundary—not because the next machine is available.
No motion outside released authority; defined stop behavior works under uncertainty, communications loss, boundary conflict, and human entry.
Independent survey/measurement demonstrates the physical task remains inside the declared tolerance envelope.
Useful work can be repeated across multiple task instances without hidden manual reconstruction of state between runs.
Interrupted work can be stopped, inspected, reconciled, and resumed or replanned without losing the authoritative task/evidence chain.
A reviewer can reconstruct what was requested, what state the system relied on, what authority was granted, what occurred, and why the result was accepted or rejected.
A downstream machine or human process can consume the accepted result without trusting an unverified completion flag.
Once one bounded earthworks cell can preserve authority, state, evidence, and recovery, Armature can add the next machine without changing what “done” means.
Back to development →