RakuAI for Robotics — The Persistent Spatial-Memory Layer Your Robots Query in Plain English

Where we sit in your stack.

ABOVE perception (we consume images, depth, and poses), BELOW high-level task planning (we serve queries to your planner's LLM). We are the scene-memory + scene-QA layer — nothing more, nothing less. That focus is the pitch.

What RakuAI provides

A per-facility Gaussian splat captured once and hosted forever
17 native MCP tools any LLM can call — "what's blocking aisle 7?", "did the pallet move since Tuesday?"
Plain-English scene QA over the reconstructed scene
Multi-vendor LLM: Claude, GPT, Gemini, or NVIDIA NIM-hosted open weights — your choice, per workload
Hosted reconstruction + query infrastructure, zero infra burden on you

What we explicitly don't build

SLAM — Cartographer, RTAB-Map, ORB-SLAM3, NVIDIA cuVSLAM already exist and are good
Motion planning — MoveIt, OMPL, Nav2 are solved problems
Robot hardware or actuation — we are software-only
Sim engines — Isaac Sim, Gazebo, MuJoCo, AWS RoboMaker; we sit alongside, not against
Action-policy foundation models — Skild, Physical Intelligence, Covariant own that layer

The wedge in one sentence: a per-facility splat that lives forever + an MCP API any LLM can query — sold per-capture and per-query, hosted, no infra burden on the customer. The moat isn't the splat — it's that the LLM is not vendor-locked, which matters to robot CTOs who already have an LLM relationship and don't want to be told which one to use.

A robot queries its facility, in one diagram.

Your robot's planner (or its LLM) talks to RakuAI's MCP surface over HTTPS — or via a stdio bridge from an on-robot Jetson Orin client. The runtime answers in spatial primitives: anchors, poses, scene queries, splat handles, object deltas.

Both halves are real and shipping. The production MCP server exposes 17 native tools (5 read-only + 12 mutation) — see engine.html for the full inventory and mcp.html for the documented scene tools. The hosted MCP runs on Azure Container Apps in East US. The LLM is yours to pick — that vendor-neutrality is the moat.

Who this is for.

Four segments, one shared pain: robots collect terabytes of spatial data that get dumped into storage and never queried again. Lead with warehouse for revenue stability; humanoid for visibility.

Warehouse automation

Robots already have SLAM. What they lack is a clean way to answer "what changed in zone B since last shift?" without writing a custom analytics pipeline. Per-site splat unifies QA across the fleet.

Example targetsSymbotic, Locus Robotics, Fetch / Zebra, GreyOrange, Geek+

Humanoid robots

Series B–D startups racing to ship a Fortune 500 pilot in 6–9 months. They buy anything that removes a 6-month engineering line item without locking them into one vendor's LLM stack.

Example targetsFigure, 1X Technologies, Apptronik, Agility Robotics, Unitree

Delivery robots

Sidewalk and last-mile fleets operate across changing urban environments. Persistent scene memory of routes, drop points, and obstacles complements existing navigation without re-instrumenting the robot.

Example targetsStarship Technologies, Serve Robotics

Inspection robots

Spot and quadruped fleets generate massive datasets dumped to storage and never opened. Customers pay per-site for auditable, time-stamped, queryable assets: "show me every valve flagged red in Q3."

Example targetsBoston Dynamics Spot ecosystem, ANYbotics, Cognite

The partner-resource stack behind a pilot.

RakuAI's NVIDIA Inception membership (accepted May 2026) and pending startup-credit programs compound here. Each is honestly labelled by status — we don't claim approvals we don't have.

NVIDIA Isaac platform

Sim · Lab · Perceptor · Manipulator · ROS

Complementary, not competitive. Isaac is what your robot sees this second; RakuAI is what your robot remembers about this facility.
Co-positioning through Inception channels removes the "isn't this what NVIDIA already does?" objection.

NVIDIA Cosmos

World Foundation Models

Capture a real facility with RakuAI → generate domain variations with Cosmos → train perception in Isaac Lab → deploy with persistent RakuAI memory.
We own steps 1 and 4; NVIDIA owns 2 and 3 — you don't glue them together.

NVIDIA Omniverse

USD digital-twin simulation

RakuAI splats export into Omniverse for digital-twin scenarios.
Your splat capture IS your digital twin — no separate CAD modeling sprint. USD export is a near-term lane.

NVIDIA NIM

Inference Microservices

Your robot's LLM can be NIM-hosted Llama / Mistral / Mixtral or Claude / GPT / Gemini direct.
MCP is the abstraction layer — privacy-sensitive on-prem NIM, high-reasoning cloud Claude, you pick per workload.

NVIDIA Brev & Jetson Orin

Inception benefits

Brev GPU cloud as a third concurrent training pool when AWS G5 / Nebius H100 are saturated.
Jetson Orin Inception hardware discount makes on-robot MCP-client eval hardware cheap to procure during a pilot.

NVIDIA DLI & Innovation Lab

Training & engineering office hours

DLI Inception code unlocks 50% off — co-train your team on Isaac + RakuAI integration.
Innovation Lab spatial-AI office hours as a co-development resource. Application status: applying — not yet granted.

AWS RoboMaker & GPU

Cloud sim · G5 / G6e · Greengrass

RoboMaker Your sim environment IS your scanned facility — close the sim-to-real gap with the same asset.
G5 / G6e for backend recon + perception inference (AWS Activate credit under review).
Greengrass our MCP client can ship as a component for fleet deployment.

Honest credit status

Where things actually stand

NVIDIA Inception accepted May 2026.
AWS Activate $10K under 5–7 day review.
Nebius H100 ($150K savings) submitted; not needed for v1.
Anthropic Claude Startups submitted, not approved — Claude is our highest-quality menu option today, not a co-marketed claim.

The ROI math.

We're not replacing your perception stack. We're replacing the persistent-memory subsystem nobody wants to own and the scene-QA LLM bridge that gets pushed to "next quarter" forever.

Status quo — build it in-house

1–2 perception engineers spend 6–18 months building persistent scene memory.
Loaded cost: $250K–$400K per engineer-year.
6-month single-engineer effort: ~$125K–$200K. 18-month two-engineer effort: ~$750K–$1.2M.
Then maintain it forever — ~20–30% of original effort per year.
Plus coordination with the LLM team to expose memory as a queryable API.

With RakuAI

$0 engineer-months.
Pay per-capture: target $0.50–$2.00 per splat reconstruction.
Pay per-MCP-query: cents per query (LLM passthrough + small markup).
Hosting included.
Your perception team ships higher-leverage work instead.

Net savings, year 1: $250K–$1M in engineering cost displaced. Honest caveat: if a prospect already has a happy persistent-memory + LLM-bridge stack, we don't have a sale — and we won't pretend otherwise.

The 90-day free pilot.

Free for the full 90 days. The output is a published, co-marketed case study — the deliverable that makes pilots #2, #3, and #4 easier to land.

Capture & connect

Weeks 1–2

Free splat capture (we scan, or coach your team) plus MCP relay live for 5–10 of your facilities. Zero invoice during this window.

Integrate

Weeks 3–8

Integration support. NVIDIA Innovation Lab office hours (once approved) booked weekly with your perception engineer. We sit in your Slack.

ROI report & case study

Weeks 9–12

Which queries got asked, how your robots used the scene memory, engineer-hours saved, what broke, what worked — then a joint published case study.

Conversion offer: per-facility license OR per-capture + per-query metering — we let the first 3 pilots tell us which pricing model wins. The pilot stays free regardless.

Apply for a pilot slot → Request a technical deep-dive →