One sentence. The AI decides which robots do it, and how.
AIRobot turns a plain-language command into coordinated action across a fleet of real robots — picking the right rig, planning the motion, running jobs in parallel, and knowing its own limits. Ten scenarios are recorded and repeatable on the bench today.
The hardest, most-doubted claim in AI robotics is no longer a promise
That an AI can run a mixed fleet of real robots from natural language — safely, and knowing its own limits — is now a recorded, repeatable demo on real hardware. What remains is adding capabilities: end-effectors, sensors, more robots. Not proving the concept.
Straight talk: today's robots have drive, arm, and sensors — but no cleaning end-effectors yet. Chore commands currently run as motion choreography (“mimed”), and the AI says so out loud. Real scrubbing needs a Phase-2 tool: a brush, mop, and fluid system. This site shows exactly what's proven today and what's honestly still ahead.
Ten scenarios. Two different robots. One plain-language brain.
Every scenario below ran live and completed on the bench, across two deliberately different robots — P1 (wireless, 6-DOF arm, compass) and MarkKOS (wired, drive-only). The operator never names a robot or lists steps; the AI does the deciding.
Verbatim transcripts from the validation runs. Replaying A1 → A5 → A8.
Routes by what a robot has
"Get a robot with an arm to move…" → AI selects P1, the only rig with an arm.
Routes by what a robot lacks
"Drive backward for a robot with no arm" → AI picks MarkKOS, reasoning from a negative capability.
A chore as real motion
"Clean the toilet" → P1 runs a 9-step arm choreography on the actual arm.
Place-aware tasking
"…in the guest bathroom" → AI resolves the room, drives there, then performs the chore.
Two chores, two rooms, two robots — concurrent flagship
One sentence → both robots navigate to different rooms at once, then perform distinct chores at once.
Proactive readiness checks
Unprompted, the AI flags low battery, tether limits, and flaky sensors — then adjusts or declines.
Capability honesty
For chores it can only mime, it says so and names the exact Phase-2 hardware required.
Watch the robots in action
Bench footage of P1 and MarkKOS running these commands live — voice in, robots moving — is being filmed now.
From a spoken sentence to coordinated robot motion
The real pipeline running on Prototype #1 today — voice in, plan, dispatch, act, report — with safety gates at every stage.
You speak — no app, no robot name
Say "robot", it answers "Yes?", and your next sentence is the command. Speech is transcribed on-device with offline Whisper — no cloud, no API key for the voice step.
Claude plans & picks the rig
The transcript goes to Claude via the Anthropic API. It reads a live capability manifest of every connected robot — battery, sensors, arm, skills, reachable rooms — then chooses the right rig and emits a step-by-step plan.
Dispatcher translates to primitives
Each step expands into controller commands — a drive uses calibrated ticksPerFoot, a skill unfolds its choreography, a navigate plays back a taught room route.
The robot acts
Commands travel laptop → COM port → wireless XBee link (or wired serial) → the on-bot Serializer controller → motors, servos, and sensors.
Sensors report back
Telemetry returns over the same path; the AI confirms the robot actually stopped, reports completion, and can auto-return the robot home by reversing the route.
Safety wins, always
Every motion is gated by an ARM check, a STOP ALL button always overrides, motions use a confirmed-stop, and robots respect a physical tether limit.
One brain, many bodies
The Cockpit console coordinates 1..N robots at once, each defined by a simple JSON file. Two rigs run on the bench today; three more are designed and gated behind a commit checklist before any hardware is bought.
4WD Mecanum chassis with a 6-DOF arm and full sensor suite. Drive, encoders, sensors, sonar, compass, arm, and servo power all validated.
A second 4WD robot on a wired serial link, drive-validated end-to-end through the Cockpit — a deliberately different platform that proves the brain is hardware-agnostic.
Bringing the 6-DOF arm fully online through P1's existing servo channels. Parts in hand; software pattern underway.
More capable platforms for real cleaning — Linux SBC, depth camera + LiDAR, cobot arm, and a fluid system. Gated behind a 6-milestone commit checklist before spend.
Built like real software, not a demo reel
The hard part isn't any single robot — it's the layer that understands intent and coordinates the fleet reliably. That's the core of AIRobot.
AI brain — Claude
Anthropic Claude (Opus) via the Anthropic API, using the Agent SDK pattern. The same robot tools are exposed over MCP so any Claude client can drive the fleet.
Capability-driven HAL
App code targets abstract interfaces (IRig, IMotor, ICompass, IServo…), so one brain runs different robots by swapping only the driver layer.
Skills are data, not code
A skill like clean the toilet is a JSON motion file. Drop in a file and any robot with the required capabilities is offered it — no code change.
Offline voice
OpenAI Whisper (tiny.en) runs locally on the laptop CPU — hands-free wake-word control with no cloud dependency for transcription.
Safety & supervision
ARM gate, STOP-ALL override, confirmed-stop, and tether limits — plus a structured event log, hang-detector, and per-rig health roll-up.
Config as data
Rigs, skills, and named locations are all JSON files. A new robot, skill, or room is added by dropping in a file — no rebuild, no code change.
An honest snapshot
This is an early-stage, single-builder hardware R&D project with working bench prototypes — not a shipping product or a funded company. Here's exactly what's real.
Ten AI-fleet scenarios validated on hardware
Capability routing, place-aware tasking, concurrent multi-robot chores, group commands, perception, and capability-honesty — all recorded across two different robots (2026-06-20).
Two robots drive-validated end-to-end
P1 (full 4WD + sensors + arm) and MarkKOS (in commission), both controlled through one Cockpit process.
Cockpit console at v0.21.0
Multi-rig observation, AI Conversation tab where Claude picks the rig and plans, calibrations with audit trail, hands-free offline voice, and auto-return-home.
Arm-on-bot + central MCP coordinator
Making the AI the commander and the Cockpit the observer, with the arm fully integrated on P1.
First real (non-mimed) chore
Toilet cleaning needs Phase-2 hardware (brush + fluid system) on a more capable platform — gated behind the commit checklist before any spend.
The concept is proven.
What scales it is capability and capital.
The riskiest question — can an AI safely run a mixed fleet of real robots from natural language? — is answered and recorded. The architecture is data-driven, so a new robot or skill is a configuration file, not a rebuild. What's next is end-effectors, more rigs, and the first real cleaning chore. If you back early-stage hardware-plus-AI, let's talk.