▌ Technology ModuleTitan Core OS

The robot OS,
built model-first.

Titan Core OS is the device OS on every Titan machine — wheeled, aerial, legged — and the VLA model is a first-class entity in it, not a bolted-on inference call. Run one VLA, or compose vision with a fast-evolving LLM that sees, plans ahead, and knows the world. Shift compute between seeing and thinking as the mission demands — and when the network allows, hand the hard calls to a datacenter-scale model. When it doesn't, the full stack runs on-device.

01 · The Stack

Open SDK.
Closed safety kernel.

Six layers on the machine, one above it — each doing exactly one job. Integrators extend the capability layer through SDK surfaces; the kernel, command link, and safety floor stay Titan-owned. The line is deliberate: extend behavior without weakening deterministic control.

FLEET LAYER

Titan Command Center

off-machine

Runs the fleet: missions, dashboards, OTA, human authority. Off the machine.

fleet orchestration · RBAC · PKI
↑ ↓
LINK LAYER

Core Command Link

Carries intent and operator overrides down, telemetry and buffered replay up. Overrides refused when the link is too slow to be safe.

Titan binary protocol · mTLS · VDA 5050 translation
KERNEL

Mission Kernel

Decides the machine's mode — connected, disconnected, degraded, fault — and runs the loop: sense → localize → plan → decide → act, on a hard reflex-rate budget.

TMO state machine · BehaviorTree.CPP composition
SKILLS

Skill Registry

Loads the machine's skills — signed, swappable software capabilities. Hardware modules stay dumb; the intelligence loads here.

MappingLocalizationPlanningDecisionsManipulationMissions
MOTION

Form-Factor Adaptation Layer

Translates one motion interface into wheels, rotors, or legs — the same mission logic drives every body.

Core-W: Nav2 + MPPI · Core-A: PX4 · Core-L: OCS2 MPC
HARDWARE

Hardware Abstraction Layer

Drives real sensors and actuators — or Titan Matrix simulation — behind the same interface. Same runtime, metal or sim.

micro-ROS MCUs · sim drivers for CI
MODULES

Hardware Modules

Dumb attachments the HAL drives: signal out, commands in, zero onboard intelligence. Cheap, swappable, and fused into one picture by the brain.

ten module types — full catalog at /core-os/modules
LocomotionSensingEmittersCommsPowerHandComputeCargoToolsInteraction
LAST RESORT

Safety Executor

An independent watchdog beneath everything: heartbeats stop, motors stop. It shares nothing with the layers above — the smart stack can fail without taking the safety floor with it.

02 · Two Modes. One Runtime.

Connectivity is a luxury.
Autonomy is the baseline.

Most "autonomous" machines stop being autonomous the moment the network drops. Titan Core OS is built for the opposite assumption — full autonomy is the default, and connectivity is what gets layered on top.

Connected Mode

Command directs. Core executes.

Titan Core OS streams telemetry to Titan Command Center, accepts mission updates, and yields to remote operators when needed. Core still owns the real-time loop — perception, control, and safety run on-device. Command sets intent; Core handles the millisecond decisions.

Telemetry uplinkContinuous
Remote override latency<200ms
Mission update pushOTA, zero downtime
Real-time control loopOn-device, <10ms
Disconnected Mode

Comms denied. Mission continues.

Jammed, out of range, underground, or under canopy — Titan Core OS doesn't notice. The autonomy stack runs end-to-end on-device. The machine keeps perceiving, planning, and acting against its last known mission. When the link returns, it resyncs telemetry and accepts new orders.

Cloud dependencyNone
Onboard map + statePersistent
Self-localizationVisual + inertial
Return-to-homeAutonomous

Graceful degradation, not graceful failure.

Cloud-tethered systems treat network loss as an exception. Titan Core OS treats it as an operating mode. There is no degraded experience when comms drop — the machine is doing the same job, on the same hardware, with the same models. The only thing that changes is whether telemetry is reaching the operator in real time. When the link returns, Core replays the buffered state to Titan Command Center and resumes connected operation.

03 · Skills

Every capability is a skill.
Swap it without forking it.

Zoom into the skill layer of the stack. Each skill sits behind a stable interface, loaded through the signed registry and metered against the runtime's latency budget. Replacing an implementation is one registration — nothing above the seam moves. Capabilities a machine doesn't carry are simply never loaded.

Signed registrationDeadline meteringAudited lifecycleLoaded only when needed

Mapping

Onboard occupancy mapping keeps a live model of the world around the machine — built on-device, carried through link loss, replayed on resync.

Localization

Confidence-aware self-localization fuses odometry with anchor corrections. When confidence decays, the runtime knows — and degrades deliberately instead of guessing.

Path Planning

Onboard planning routes toward mission goals and replans the moment the map changes — a blocked corridor is a detour, not a stop.

Model-First Decisions

The decision layer is a model slot, not a hardcoded brain: one VLA, or vision composed with a fast-evolving LLM. Priority shifts with the mission — more compute for seeing or for thinking — and when the link allows, a datacenter-scale model takes the hard calls. Low-confidence decisions escalate to a named operator.

Manipulation

The hand is a hardware module with near-zero intelligence of its own. Mount it and the brain loads the grasp skill — force limits, sequencing, audited actuation. Remove it and nothing else changes.

Mission Profiles

Patrol & Security, Delivery & Logistics, Agriculture & Field Work, Search & Rescue — each vertical is a mission profile the planner consumes. A new vertical is a new skill, not a new runtime.

Optional Redundancy

When one of anything isn't enough.

Any capability slot can hold more than one provider — a primary localizer with a hot standby, a primary planner with a conservative fallback. An arbiter watches health, confidence, and deadline behavior, and fails over the moment the active provider degrades — then hands back when it recovers.

Opting inOne registration, no runtime changes
Single-provider machinesZero overhead
Every switchoverLogged to the audit chain
Safety fallback layerIndependent — never arbitrated
04 · Substrate & Builds

ROS 2 substrate.
Titan-owned runtime.

Titan Core OS uses ROS 2 Jazzy as the driver and message substrate, with rmw_zenoh as the default RMW and Cyclone DDS available for legacy compatibility. Peripheral MCUs connect through micro-ROS over XRCE-DDS or Zenoh-pico.

The substrate stays open enough for OEM integrators and defense programs. The runtime contracts stay strict enough to bound mode transitions, actuator authority, plugin behavior, and OTA rollback.

HAL drivers swap real sensors and actuators for sim drivers inside Titan Matrix, so the same runtime can run on hardware and in CI without recompiling the autonomy stack.

What it runs on
ComputeNVIDIA Jetson Thor, Orin, or x86 + RTX dev targetsv1
InferenceTensorRT + ONNX Runtime; optional Hailo companion for always-on perceptionv1
DefaultPREEMPT-RT Linux, Ubuntu 24.04 LTS Prov1
Safety partitionQNX 8.0 SDP for ASIL-D / MIL-STD programsv1.1
Mixed criticalityXen with PREEMPT-RT domU and QNX safety domainv1.1

One runtime. Three specializations.

The shared layers — perception, planning, autonomy fallback, security — are identical across builds. Only the motion layer of the stack changes.

Core-W

Titan 0000

Wheeled

Greenfield rate-loop and steering controller with Nav2 + MPPI planning for off-road, yard, closed-site, and defense mobility.

Core-A

Titan 88

Aerial

Titan attitude and position controller with PX4 low-level flight firmware over uXRCE-DDS and wind-aware FFAL models.

Core-L

Titan 1111

Legged

OCS2-based convex MPC and whole-body-control fork with Titan terrain perception and gait selection above the controller.

05 · Why It Matters

Latency kills.
Literally.

At 60 km/h, a 100ms cloud round-trip means 1.7 meters of blind driving. At 120 km/h, it's 3.3 meters. In defense scenarios with adversarial jamming, the cloud doesn't exist at all.

Titan Core OS's on-device inference is a structural advantage in both reliability and security. It isn't a technical preference — it's an architectural requirement for any system where latency variance can cause harm.

Core guarantees bounded, deterministic latency under all conditions — full connectivity, degraded connectivity, or zero connectivity. The same deterministic on-device.

Titan Core OS<10ms
On-device, deterministic, zero network dependency
Cloud (Best Case)50–80ms
Fiber, nearby region, no congestion
Cloud (Typical)100–200ms
Real-world conditions, variable tail latency
Cloud (Degraded)500ms+
Congestion, distance, interference, or denial
06 · Trust & Authority

Defense-grade is
the default build.

The v1 runtime assumes regulated fleets, contested networks, and named human authority over every actuator command. Security and audit are not add-ons; they are part of boot, update, and control.

TPM-backed identity

Device certificates live behind TPM 2.0, are managed by Titan Command Center PKI, and support Titan-issued or customer-owned roots.

Signed plugin boundary

Production loads only Titan-signed or customer-signed plugins. Unsigned modules run only in dev mode with sandbox warnings.

Native evidence trail

MCAP recordings, crash dumps, OpenTelemetry traces, and actuator attribution ship to Command when the link returns.

07 · Resync

When the link returns,
Core catches Command up.

Titan Core OS buffers telemetry, decisions, sensor highlights, and mission state during disconnected operation. The moment a link is re-established with Titan Command Center, Core replays the buffer at high-priority — operators see the gap close, edge cases get flagged for review, and any new orders take effect immediately.

Every disconnected mission is also a training-data event. Buffered edge cases feed the data flywheel — the fleet learns from the moments when the network wasn't watching.

08 · Resilience Loop

Tested in worlds
that don't exist.

Titan Core OS runs continuously inside procedurally generated worlds — physics-accurate scenes from game engines, and unrealistic environments from video generation models. We deliberately corrupt sensors, fail actuators, and break physics. The OS endures, or we fix it. Then we do it again.

Game-engine worlds

Physics-accurate. Fully controllable.

Procedurally generated terrain, weather, traffic, lighting, and sensor noise. Every variable is dialable — the right tool for systematic stress tests and exhaustive edge-case enumeration.

Video-model worlds

Surreal. Novel. Unhinted-at by training.

Generative video models produce scenes our engineers wouldn't think to script — broken physics, unfamiliar objects, impossible lighting. Core has to make sense of worlds it has never seen and would never see in reality.

01

Generate

Game engines and video models continuously spawn new worlds — physics-accurate or deliberately surreal.

02

Run

Titan Core OS boots inside the world and runs the full autonomy stack — perceive, plan, act, recover.

03

Observe

Decisions, latencies, recoveries, and crashes are logged. Did the OS endure, fail safe, or fail loud?

04

Harden

Failures feed model training and runtime fixes. Validated changes ship via OTA. Loop repeats.

Failure modes we deliberately inject

The runtime should survive reality. Reality doesn't ask permission, so neither do we.

Sensor degradation
LiDAR dropouts, camera blooms, IMU drift
Actuator failure
Motor stalls, partial loss of degrees of freedom
Comms denial
Jamming, intermittent uplink, packet corruption
Compute saturation
Thermal throttle, OOM, single-bit flips
Adversarial inputs
Sensor spoofing, GPS denial, visual occlusion
Impossible physics
Surreal video-model worlds with broken constraints

The pact: when Core can't endure, it must fail safe — degrade gracefully, hand back control, return home, or shut down without harm. Crashing isn't a failure mode we accept. Every crash in simulation is a runtime fix we ship before any robot encounters it in the field.