Lupotek

SUBTERRA

NAVIGATE BEYOND
THE SIGNAL

Nomadic spatial intelligence for autonomous machines in GPS-denied underground environments.

Subterra establishes a spatial reference inside the mission itself. Our onboard architecture jointly estimates robotic motion, reconstructs unfamiliar environments and maintains the uncertainty attached to every consequential spatial decision — without depending on GNSS, cloud connectivity or a surveyed positioning network.

Multi-modal simultaneous localisation and mapping (SLAM) integrates geometric, inertial and semantic evidence into a persistent world model. Observability, measurement provenance and navigation integrity govern how that model informs machine action.

Designed for direct integration with autonomous robotics, Subterra supplies pose, covariance, environmental structure and admissible motion corridors to wheeled, tracked, legged and aerial platforms. Spatial continuity extends across sensing conditions and mobility domains while each machine retains its own operating envelope.

The coordinate system travels with the mission. Intelligence operates on the machine.

01 / SPATIAL INDEPENDENCE

Global Needs

Autonomous capability where conventional positioning ends.

The Challenge:

Underground machines must understand where they are, what surrounds them and where they can move — without relying on infrastructure that may not exist.

Mines, tunnels, shafts, caves and subterranean infrastructure combine weak communications, difficult terrain and unstable perception. Darkness limits cameras; dust and airborne water degrade optical measurements; repetitive corridors obscure position along the tunnel axis. Dead reckoning accumulates drift, while a convincing but incorrect place match can distort an entire map.

The Solution:

Subterra creates a mission-local spatial reference and jointly estimates the machine’s trajectory and environment. It couples navigation confidence to autonomous behaviour, so the system can seek stronger observations, reduce speed, recover a known location or return when the available evidence no longer supports continued exploration.

Our architecture addresses:

  • GPS-denied operation without a pre-surveyed positioning network;
  • Unknown terrain and infrastructure with no prior map;
  • Perception degradation across darkness, dust, smoke and moisture;
  • Geometric degeneracy and repeated tunnel structures;
  • Accumulating drift and unreliable loop-closure candidates;
  • Disconnected missions and intermittent robot-to-robot communications;
  • Shared mapping across distinct robotic mobility domains;
  • Direct delivery of spatial intelligence to robot planners and controllers.
02 / NOMADIC ARCHITECTURE

Spatial Intelligence

Pose, geometry, evidence and admissible action — maintained in one probabilistic spatial context.

S∷
SUBTERRASPATIAL INTELLIGENCE / NOMADIC CORE
ST—04SUBSURFACE / MULTI-PLATFORM MISSION
MISSION FRAMES0Portal-relative / SI units
OBSERVED CONNECTIVITY344 m10 observed / 12 configured passages
REFERENCE POSE / MAX 1σ0.052 mCONSTRAINED / configured noise
GRAPH STATELOCAL0 admitted / 2 registered platforms
Metric + semantic world modelLEG-02 · KEYFRAME 82
P00F02S01UGV-01LEG-02H01 / RUBBLE TRANSITIONXYZ / OBSERVED REFERENCE GEOMETRYS0 · 4,320 REFERENCE SAMPLESX20.0 m / 1.0×AIR-03 / INDEPENDENT FRAMES0 POSITION WITHHELDRange evidence ≠ rigid alignment
— Trajectory□ Pose / 1σ ellipse ×20△ EvidenceDotted / unobserved geometry
SELECTED VOLUME / T031.80 m / PASSAGE WIDTH0.75 m
REFERENCE BODY / T(S0,B)
Translation / m
31.70 / 4.04 / 0.00
σx / σy / σz
0.052 / 0.041 / 0.034
Acquisition context
Onboard / reference mission
k082 / reference case loaded · evidence available at cursor
Decision history
    S0 / GNSS-INDEPENDENT REFERENCEREFERENCE MISSION · LOCAL COMPUTATION
    ST–04 / ENGINEERING REFERENCE

    The mission workspace makes our architectural decisions inspectable: measurement weighting, directional uncertainty, graph admission, frame alignment, platform constraints and recoverable routes. Its configured mission inputs produce calculated outputs; they are distinct from acquired sensor telemetry and measured field performance.

    The reference case evaluates a conditional pose projection, a translation graph and a passage-level planner. Deployment acceptance extends to the complete velocity–bias estimator, calibrated sensor timing, volumetric collision checking, platform control and independently measured operating performance.

    LOCAL REFERENCE

    A coordinate system created in motion

    Subterra initialises an internally consistent mission frame from the robot’s starting pose. Subsequent observations constrain six-degree-of-freedom motion relative to that frame, without requiring latitude, longitude, an external heading reference or a satellite-derived clock.

    Local consistency and global georeferencing remain distinct. Where an authorised surveyed tie becomes available, its transform can align the mission map to an external frame while retaining the alignment uncertainty.

    ROBOT-NATIVE WORLD MODEL

    More than a position estimate

    Our spatial state combines pose, velocity, sensor biases, geometric maps, a SLAM factor graph, covariance, semantic entities, peer-agent relationships and navigation integrity.

    Metric geometry, connectivity, semantic evidence and platform-specific traversability describe the same environment at complementary resolutions. Free, occupied and unobserved space remain distinguishable; hazard hypotheses retain their supporting observations rather than becoming unqualified map facts.

    01Multi-Modal SLAM & Factor-Graph Estimation

    Our estimator jointly resolves trajectory and map from environmental and motion observations. Pose, velocity and bias states form variable nodes; IMU preintegration, LiDAR registration, radar velocity, visual landmarks, kinematics, relative ranges and validated loop closures contribute uncertainty-weighted factors.

    Measurement admission combines residual consistency, sensor health, timing and calibration. The state covariance retains directional uncertainty and cross-state coupling; robust weighting limits inconsistent factors. A bounded local estimation window supports onboard execution, while long-range constraints reconcile accumulated trajectory and map drift.

    02LiDAR-Inertial Geometry

    Sequential 3D point clouds constrain rigid-body motion through scan-to-scan and scan-to-map registration. Inertial measurements provide high-rate motion priors and support point-cloud motion compensation during a scan.

    Walls, floors, ceilings, rails, pipes and shaft geometry contribute structural constraints. Subterra evaluates the strength of those constraints independently across translation and rotation, rather than treating a successful registration as proof that every degree of freedom is observable.

    03Radar-Inertial Resilience

    Radar contributes range, angular structure and Doppler observations when optical sensing deteriorates. Static-return selection and outlier rejection allow radial velocities to constrain the platform’s ego-motion alongside inertial measurements.

    Our fusion architecture changes the influence of each sensing domain according to measurement quality. Radar complements optical geometry; multipath, specular surfaces, moving objects and weak returns remain explicit sources of uncertainty rather than being concealed by a single confidence score.

    04Visual, Thermal & Semantic Perception

    Visible-light and thermal observations add feature tracking, place recognition and operational context where the environment supports them. Depth and inertial priors can strengthen visual estimation under changing illumination or limited texture.

    Semantic entities — including junctions, doors, infrastructure, people, standing water and obstructions — are spatially registered with their observation history and confidence. A suspected hazard remains distinguishable from a verified structural observation.

    05Cross-Modal Place Recognition

    Subterra’s place-recognition architecture relates geometric, visual, thermal and radar representations to a common spatial context. A location observed through one sensing domain can become a candidate match when another domain is dominant on a later visit.

    Learned similarity proposes associations; geometric consistency, available independent observations and uncertainty checks determine whether a match is admitted to the spatial graph. Repeated tunnel segments do not become trusted loop closures on appearance alone.

    03 / OPERATING MODEL

    How Subterra Works

    A continuous estimation–decision loop, governed by evidence, uncertainty and recoverability.

    Subterra links perception, state estimation and autonomy in a closed operational loop. Mapping informs motion; motion creates new evidence; new evidence revises the world model and the next permissible action.

    Our nomadic AI system draws on learnings from the LTKAU Initiative’s Lutine AI system, particularly evidence handling, semantic relationships, uncertainty-aware reasoning and bounded agent objectives. The learning layer proposes associations and mission objectives; probabilistic estimation admits spatial constraints, and deterministic supervision governs motion authority.

    01Initialise — Establish the Mission Frame

    Define the starting pose, initialise inertial state and sensor transforms, and establish the local frame and mission clock. Confirm the selected platform’s kinematics, sensing configuration, energy constraints and permitted operating envelope.

    No fixed anchors or existing map are required to begin relative localisation. Absolute position and heading are introduced only when supported by an external reference.

    02Observe — Evaluate the Available Evidence

    Synchronise sensor streams, compensate motion, reject invalid observations and maintain calibration state. Assess image quality, geometric structure, radar consistency and kinematic reliability before assigning measurement influence.

    Wheel slip, uncertain footholds, dust contamination and delayed peer observations can change the evidence available to the estimator without changing the mission’s underlying spatial frame.

    03Estimate — Maintain Pose, Map & Uncertainty

    Update six-degree-of-freedom pose and motion within the local reference. Fuse valid constraints into the evolving graph and retain covariance with the state, rather than publishing an unqualified coordinate.

    Keep a locally continuous odometry frame for robot control and a globally corrected map frame for long-horizon planning. Loop-closure corrections update the relationship between those frames without silently imposing a discontinuous command on the controller.

    04Detect Degeneracy — Identify Weakly Constrained Motion

    Evaluate the local information structure and registration residuals to identify directions with insufficient geometric support. A feature-poor corridor may constrain lateral displacement well while leaving longitudinal motion uncertain.

    Subterra increases uncertainty along weak directions and seeks complementary measurements. When useful evidence remains insufficient, the navigation integrity state reflects that loss of observability.

    05Plan Actively — Move to Improve Understanding

    Active SLAM balances mission progress against terrain risk, energy, uncertainty and expected information gain. The next useful observation may require a different viewpoint, a junction scan, a return to known geometry or a cooperative rendezvous.

    The preferred path is evaluated for the specific robot. A shorter route can be rejected when it offers poor localisation, inadequate clearance or an unacceptable loss of return capability.

    06Close Loops — Reconcile Drift

    Search for previously observed places, verify candidate correspondences and add defensible long-range constraints. Graph optimisation then corrects the affected trajectory and spatial model.

    Outlier-resistant optimisation and consistency checks reduce the risk of a false match collapsing separate tunnel segments into one. Corrections and their supporting observations remain traceable.

    07Cooperate — Align Independently Created Maps

    Each robot can explore in its own reference frame. Overlapping geometry, relative pose or sufficient ranging observations constrain the transformation between frames and support collaborative map alignment.

    A single peer range does not establish a complete frame transformation. Subterra retains the unresolved degrees of freedom until additional geometry or measurements support alignment, and accounts for shared information when combining estimates.

    08Continue Offline — Recover & Synchronise

    Local estimation, mapping and permitted navigation execute onboard during a communications outage. Mission evidence and map updates are retained for exchange when a usable link returns.

    Peer observations carry timestamps and frame identifiers; stale information is not treated as current. Return, hold and recovery behaviours remain constrained by localisation integrity, energy reserve, terrain and the approved mission policy.

    04 / CROSS DOMAIN

    Multi-Domain Mobility

    A shared spatial reference. Distinct kinematics, traversability models and control contracts.

    Subterra separates spatial intelligence from platform morphology. Robot-specific adapters connect the common world model to the sensors, kinematic constraints, planners and control interfaces of each machine.

    A ground platform can map a primary tunnel, a quadruped can extend exploration across broken terrain, and an aerial platform can survey a shaft or upper chamber. Validated frame alignment preserves spatial continuity while each robot retains its own mobility and safety constraints.

    ONE SPATIAL COREROBOT-NATIVE EXECUTION / DISTINCT MOBILITY ENVELOPES
    S0 / SHARED MISSION CONTEXT
    CONTROL CONTRACTBounded twist
    BODY CLEARANCE1.20 m
    SPATIAL FRAMES0 / shared mission

    Primary tunnel geometry is retained when rubble exceeds the tracked mobility envelope. The obstruction, observation provenance and recovery corridor remain available for a legged handover.

    COMMON MAP / PLATFORM-SPECIFIC POLICYMISSION EVIDENCE / RETAINED
    01Wheeled, Tracked & Industrial Platforms

    Traversability combines gradient, surface roughness, clearance, traction, turning radius and vehicle dimensions. Wheel and track motion contribute odometry where reliable, while slip estimates reduce their influence when ground conditions change.

    Applications include autonomous haul vehicles, tunnel survey robots, inspection crawlers, rail platforms and suitable drilling or support systems.

    02Legged & Climbing Systems

    Kinematic observations and contact state supplement environmental sensing. Planning considers foothold stability, step height, body clearance, contact uncertainty and the distinction between a geometrically reachable surface and a physically supportable one.

    Rubble, stairs and irregular transitions can be assessed against the capabilities of a specific quadruped or climbing platform without assuming that all ground robots share the same traversability envelope.

    03Subterranean Aerial & Hybrid Platforms

    Aerial planning uses free-space volume, rotor clearance, airflow constraints, energy and localisation quality. Narrow shafts and overhangs require a different collision and recoverability model from ground navigation.

    Ground-carried aerial platforms can extend a mission into areas the carrier cannot access. Shared geometry and launch context support map alignment, while the aircraft maintains its own onboard estimation and control loop.

    04Cooperative Navigation & Mobile References

    Robots can exchange spatial constraints through relative visual pose, LiDAR or radar observations, and suitable peer-ranging systems such as UWB. Local radio or acoustic measurements support cooperation without requiring internet connectivity.

    Optional breadcrumb nodes can provide ranging, communications relay or homing cues. Their deployment pose is initialised from the robot’s estimate, including its uncertainty; they are jointly refined as further observations arrive. Self-deployed nodes supplement the architecture rather than becoming a pre-installed prerequisite.

    05Direct Integration Into Autonomous Robotics

    Subterra’s integration architecture exposes timestamped pose, velocity, transform trees, covariance, point clouds, occupancy or voxel maps, semantic entities, dynamic obstacles, traversability, trajectory history and navigation integrity to the host robotics stack.

    ROS 2-compatible adapters and platform APIs carry frame identifiers, acquisition timestamps, covariance and integrity with each output. Each contract specifies transform direction, covariance ordering and perturbation convention, clock domain, update rate, latency budget, quality of service and stale-data handling. Sensor extrinsics and timing offsets belong to the estimation and calibration contract; a delayed observation cannot silently acquire the time or frame of its arrival.

    Map-frame corrections remain separate from the continuous odometry consumed by the controller. A correction invalidates affected route assumptions and triggers bounded replanning; it does not become an instantaneous body command. Spatial outputs inform autonomy while the host platform’s deterministic controllers enforce motion limits, stopping behaviour and recovery authority.

    05 / STRATEGIC VALUE

    Subterra Advantage

    Infrastructure-independent spatial intelligence with inspectable decisions and bounded autonomy.

    Subterra connects perception to machine action through an explicit spatial and evidential model. A registered obstruction carries geometry, confidence, acquisition history and a platform-specific mobility consequence; a proposed route carries its clearance, uncertainty, energy and recovery constraints.

    Learnings from Lutine inform the relationship between observations, inferred conditions, dependencies and permitted responses. That continuity allows a changed passage, degraded sensor or unresolved peer frame to revise mission decisions without discarding the evidence that produced them.

    We develop Subterra around measurable navigation integrity, repeatable testing and platform-specific acceptance. Local accuracy, long-term consistency, degraded-mode behaviour and mission completion are evaluated separately.

    01Infrastructure Independence

    Enter unmapped environments without requiring GNSS, cellular positioning, Wi-Fi fingerprints, fixed beacons or cloud inference. Onboard compute and the selected sensor payload establish the mission’s relative spatial reference.

    Infrastructure independence does not remove drift or guarantee observability. The architecture explicitly manages the uncertainty that accumulates when useful spatial constraints are absent.

    02Navigation Integrity & Controlled Autonomy

    Publish estimate age, covariance, sensor health, observability and integrity status alongside position. Policy-defined limits determine when the robot may continue, seek better evidence, slow, hold or attempt recovery.

    Operational authority is expressed through mission scope, permitted actions, energy and evidence budgets, and escalation conditions. Safety-critical motion constraints remain enforceable when communications are unavailable.

    03Persistent Semantic & Spatial Memory

    Retain geometry, topology, object identities, observations and hazard hypotheses across mission sessions. Relocalisation reconnects a returning platform to previously established structure when the evidence supports it.

    Changed infrastructure and uncertain associations are represented explicitly. Older maps provide context without automatically overriding current observations.

    04Validation & Assurance

    Acceptance separates geometric accuracy, statistical consistency and operational integrity. A locally plausible trajectory is insufficient: relative and absolute trajectory error, map distortion, uncertainty coverage, false-loop acceptance, relocalisation and recovery must be evaluated against independent ground truth and the declared mission frame. A small covariance is credible only when its calibration survives those comparisons.

    Each operating envelope requires repeatable sensor-dropout, dust, low-light, repetitive-geometry, slip, clock-offset and calibration-perturbation trials. Cross-platform tests challenge collinear or insufficient alignment evidence, stale observations and duplicated information. Communications-loss tests verify that onboard decisions and return policies remain enforceable without peer exchange.

    Compute acceptance includes tail latency, deadline misses, memory growth and thermal or power limits on the selected onboard hardware. Hardware-in-the-loop and controlled field trials must establish stopping, retreat and map-correction behaviour before a deployment envelope is accepted. Technical due diligence distinguishes reference calculations, recorded-data results and witnessed platform trials; performance claims require the corresponding evidence.

    05Architecture Boundaries & Evidence Ownership

    Subterra separates geometric estimation, learned perception, mission reasoning and motion supervision. Learned outputs carry observation provenance and uncertainty; they do not replace geometric verification or obtain unrestricted controller authority. An unavailable inference model must degrade semantic capability without invalidating otherwise supported local odometry.

    Peer alignment is admitted only when the available geometry constrains a proper rigid transformation. Its covariance, timestamp and source relationships accompany the transformed evidence. Shared observations are tracked to prevent confidence being inflated by repeated exchange of the same information.

    A navigation decision records its frame, acquisition context, applicable platform envelope and reason for admission or withholding. Unknown terrain, unresolved references and stale evidence retain their meaning through planning, export and review.

    06Mining & Critical Infrastructure

    Support underground survey, inspection, autonomous logistics and condition awareness across mines, utility tunnels, rail infrastructure and industrial voids. A shared spatial model connects infrastructure entities, route constraints and robot-accessible work areas.

    07Emergency Response & Hazardous Environments

    Support reconnaissance and search in dark, obstructed or communications-degraded spaces where direct human access is hazardous. Spatially registered observations help teams understand access routes, suspected hazards and the limits of available coverage.

    08Research & Extreme-Environment Exploration

    Provide a common architecture for cave, subsurface and planetary-analogue exploration, with mission profiles adapted to platform constraints and the available sensing domains. Preserve local spatial understanding where external positioning cannot be assumed.

    09Australian-Controlled Development

    Build around modular onboard software, controlled model updates and customer-defined data custody. Sensor configurations, robot interfaces and deployment environments are selected against the operational requirement, preserving flexibility across platform suppliers.

    06 / PARTNERSHIPS

    Engage with Subterra

    Subterra is seeking strategic collaboration with government, mining, critical-infrastructure, robotics, research and technology partners. Qualified organisations may request an initial briefing, technical information or access to Lupotek’s controlled engagement and due-diligence process.

    Contact For More Information