Robotics R&D Pilot Stack

MANTIS is our embodied‑AI pilot stack.

A live RobotOps cockpit and validation framework for Jetson-powered robotics — connecting vision, speech, memory, service health, safety gates, and supervised action into one operator surface.

SSE LiveObservation-OnlyJetson Orin Memory FabricQA TripwiresROS2 Bridges

MANTIS is where we integrate the hard parts of embodied AI: perception, speech, memory, edge telemetry, supervised actions, and real-time operator visibility.

Why MANTIS

Robots do not fail in one place.

Embodied AI systems fail across surfaces — models, memory, speech, frontend contracts, containers, edge devices, ROS2 bridges, cache versions, and physical safety boundaries. MANTIS exists to make those failures visible before they reach the operator or customer.

01 / Surface

Multimodal complexity

Vision, language, speech, memory, and actions must stay synchronized.

02 / Surface

Edge reality

Jetson thermals, RAM, power, containers, and services affect behavior.

03 / Surface

Operator trust

Every proposed action needs visible state, traceability, and supervision.

04 / Surface

Demo reliability

Customer-visible regressions become reusable tripwires.

System Topology

MANTIS is not one model.

It is the orchestration and validation layer around a distributed embodied-AI stack.

Inference · Memory · Speech

MotherPC

  • LLM / Hermes
  • Riva Speech
  • Honcho Memory
  • Embed Service
  • Triton Inference
Edge · Bridges · Telemetry

MANTIS-Orin

  • VLM
  • ROS2 Bridges
  • Jetson Telemetry
  • Container Health
  • Edge Services
Operator · Scenario · Safety

CD5 Cockpit

  • Scenario Runner
  • Live Vitals
  • Safety State
  • Service Health
  • CX Canaries

The Operator Surface

CD5: the operator surface for MANTIS.

CD5 turns the robotics stack into a live control surface: service health, Orin vitals, memory state, speech state, scenario execution, safety posture, and container status in one interface.

/api/orin/vitals

Live Orin telemetry

GPU, CPU, RAM, thermals, power, fan, disk, and NVP model.

/api/services/health

Service health

MotherPC and Orin services tracked independently.

/scenario/run

Scenario runner

Run vision, memory, speech, and safety probes from one panel.

/sse/state

SSE live state

Operator view updates continuously without page refresh.

/trace/raw

Raw trace access

Debug the exact endpoint, response, and failure path.

/containers

Container visibility

Know which services are running, absent, paused, or degraded.

Repeatable Probes

Repeatable scenarios, not one-off demos.

Every scenario is a repeatable probe. MANTIS treats demos as testable flows, not theatrical one-offs.

/scenario/describe-scene

Describe Scene

Target  under 1800msRaw Trace
/scenario/count-objects

Count Objects

Target  under 1600msRaw Trace
/scenario/ask-mantis

Ask MANTIS

Target  under 2200msRaw Trace
/scenario/remember-fact

Remember a Fact

Target  under 600msRaw Trace
/scenario/recall-fact

Recall a Fact

Target  under 800msRaw Trace
/scenario/voice-loop

Voice Loop

Target  under 2400msRaw Trace
/scenario/asr-probe

ASR Probe

Target  under 700msRaw Trace
/scenario/tts-probe

TTS Probe

Target  under 900msRaw Trace
/scenario/safety-probe

Safety Probe

Target  under 200msRaw Trace

Continuity Layer

Memory is infrastructure, not a prompt trick.

Honcho, Embed, and Bus form the continuity layer behind MANTIS.

memory.honcho

Honcho

Persistent agent and user memory for goals, preferences, instructions, and context.

memory.embed

Embed

Semantic retrieval for previous observations, transcripts, scene descriptions, and task context.

memory.bus

Bus

Event flow between vision, speech, memory, UI, ROS2, and service health.

MANTIS uses memory to make robot interaction continuous across turns, sessions, users, and environments. The operator can verify when memory is used, when it is stale, and when it is unavailable.

Safety Posture

Observation-first. Action-gated. Operator-visible.

MANTIS separates perception and proposed action from physical actuation. Operators can inspect what the system saw, remembered, proposed, and blocked before any motion path is enabled.

Observation-Only

Proposed actions are read-only unless motion controllers are explicitly enabled.

  • Action proposals are visible
  • Motion is gated
  • Service state is inspectable
  • Safety posture is always shown
  • Operator supervision remains central

CX Tripwires

Green tests are not enough. The customer experience must be green.

MANTIS converts escaped demo failures into reusable canary gates. If a frontend/backend contract drifts, a stale asset ships, or one user action creates duplicate side effects, the framework rejects the lane before it closes.

Verifies · Frontend to Backend

Response Shape Parity

Verifies the fields the frontend reads actually exist in backend responses.

frontend reads payload.status. backend returns no status field.
Verifies · HTML to JS Build Tag

Cache-Bust Agreement

Verifies served HTML and JavaScript agree on the active build tag.

index.html build a1c4 / app.js build b9f2
Verifies · One Action One Effect

Duplicate Action UUID

Verifies one user action cannot trigger duplicate side effects.

action uuid 7f31 emitted twice within 800ms.

A previous voice issue passed server-side checks but failed in customer testing because frontend and backend response shapes disagreed. MANTIS now turns that class of failure into a reusable detector.

CD5.GATE.A1 — Customer-Experience Tripwire Trio

Current Status

Pilot stack under active R&D.

Stage
Pilot / R&D framework
Focus
Embodied-AI operations and validation
Hardware
MotherPC + Jetson Orin
Runtime
Multimodal services + ROS2 bridges
Safety
Observation-first, supervised action
Direction
Field-ready RobotOps framework

MANTIS is the current proving ground for our robotics architecture. The goal is not to hide complexity, but to make it observable, testable, and safe to operate.

Built For

Built first for robotics builders.

01

Robotics labs

Repeatable embodied-AI demos with live health and traceability.

02

Jetson · ROS2 teams

Edge telemetry, model services, bridges, and containers in one cockpit.

03

Field demo engineers

Know what failed before the customer does.

04

AI platform teams

Convert customer-visible failures into reusable gates.

05

Technical pilot partners

Bring a real robot stack and validate it under supervised operation.

Next

Bring your robotics stack. Make it observable, demo-ready, and failure-aware.

See it. Remember it. Monitor it. Gate it.

← Back to TOF.STUDIO