Home robots you can trust.
Arkona builds a perception-grounded safety harness for physical AI — every proposed action must be affirmatively confirmed safe by perception before it executes, checked against a structured world model and a forward-simulated trajectory, or it is blocked by default.
Proud member of theNVIDIA Inception Program
Perception · Prediction · Planning · Dexterous manipulation — unified in one learned model.
We learn how the physical world responds — then we act.
A world model is a learned, predictive simulator of the robot's environment. Instead of reacting frame by frame, the aim is for Arkona's robot to imagine the outcome of each candidate action and choose the one most likely to succeed.
This is our research direction. What's actually validated and shipped today is the safety harness below — a rule-based, perception-grounded gate that runs in front of whatever proposes an action (a learned policy, a planner, or this world model once it exists) and blocks it by default unless every precondition is confirmed.
Predict before acting
Every motion is rehearsed inside the model first. The robot commits only to plans it expects to succeed, then corrects in real time as reality diverges.
Generalises across tasks
Because the model captures physics — contact, weight, friction, occlusion — skills transfer between builds, objects and rooms instead of being hand-scripted one by one.
Fails safe, not clever
When a precondition can't be confirmed, the robot doesn't try to be clever about it — it freezes in place or retracts to a known safe pose, simple enough that the fallback itself can't inherit the same coverage gaps it's guarding against.
A safety harness validated live in NVIDIA Isaac Sim.
Arkona's perception-grounded safety harness is validated live in NVIDIA Isaac Sim and Isaac Lab: every proposed action must be affirmatively confirmed safe by perception before it executes, checked against a structured world model and a forward-simulated trajectory — with any unconfirmed condition defaulting to blocked.
Master safe manipulation in simulation and you have mastered the core of grasping, placing, and reacting safely around people — the exact skills any home task demands.
The safety pipeline, end to end
Perceive the scene
A structured WorldState is built from live sensor and proprioceptive data — objects, agents, robot state, environment signals.
Predict the swept path
A dynamics adapter forecasts exactly where the proposed action will take the robot before anything moves.
Check every precondition
31 checks run against the prediction — object and payload safety, ISO/TS 15066-referenced human-proximity limits, sensor integrity, command/config security, robot self-limits.
Decide: permit or block
Only a full PERMIT — every check satisfied — reaches the actuators.
Execute, or fall back safely
A failed check triggers a safe, pre-defined fallback instead of the proposed action.
Log and escalate
Every decision is logged for audit — the record a real safety review needs.
Four capabilities we're building toward.
This is the direction our research is aimed at — a single learned representation spanning perception, prediction, planning and control. None of this is what's validated today: today's validated system is the narrower, rule-based safety harness described above. Treat the four pillars below as roadmap, not a status report.
Multimodal perception
Fuses colour, depth and touch into a coherent 3-D understanding of the scene and the objects in it.
Predictive world model
Rolls out imagined futures so the robot can weigh actions against their likely physical outcomes.
Instruction grounding
Connects human instructions — printed steps, language, diagrams — to concrete actions in the world.
Dexterous control
Force-aware manipulation that grasps, aligns and seats parts with the precision a home demands.
Meet the testbed: Arkona P-1.
Our first-generation research cell pairs a 6-axis manipulator with overhead and wrist cameras above an instrumented work surface — the platform where the world model meets real objects.
The same safety harness that's validated live in NVIDIA Isaac Sim is what lets the robot work calmly and predictably around people — sensing its surroundings continuously and stopping the instant something unexpected enters its space.
AI-first — and safe, secure, reliable by design.
Modern AI is the foundation everything else stands on. On top of that base, three principles govern how the robot behaves in your home — not bolted on afterwards, but part of how the system perceives, predicts and acts.
AI foundation
Large multimodal models give the robot broad commonsense and language — the bedrock its perception, world model and control are built on and continually improved with.
Safety
Force-limited, collision-aware motion with hardware e-stops. The world model predicts contact before it happens, so the robot slows and stops around people and pets.
Security
Tracking people and logging every decision means this system processes personal data — a genuinely different regulatory axis from safety. Our design doc names that as an open item, not yet solved: a real deployment addresses data protection (GDPR, CCPA, retention, anonymization) deliberately before going live, not by default.
Reliability
Degraded perception — low light, occlusion, a sensor fault — lowers confidence on every downstream check rather than being ignored, so the robot fails closed instead of guessing. Repeated blocks on the same action escalate to a human rather than retrying indefinitely.
From simulation to the whole home.
Two robots validated
Franka Panda arm + ANYmal-C quadruped, validated live in NVIDIA Isaac Sim and Isaac Lab.
Real-hardware validation
Validate the full safety harness end to end on real hardware.
Humanoid support
A Unitree G1 adapter now exists and is in early integration — hazard-campaign validation against it hasn't started yet.
Everyday home tasks
Carry the same safety harness to real chores, on whatever robot proposes them.
Building the future of home robotics?
We are talking to researchers, hardware partners and early collaborators who want to help teach robots to understand the physical world.
We’re raising a pre-seed round.
Arkona builds a perception-grounded safety harness for physical AI: every proposed action must be affirmatively confirmed safe by perception, checked against a structured world model, before it executes. First proving ground: safety-harness, validated live in NVIDIA Isaac Sim and Isaac Lab.
Recognition is solved — trust isn’t
Foundation models already know a home’s commonsense. Acting on it reliably and safely, around people, is the open problem.
A perception-grounded safety harness
Every proposed action is checked against a structured world model and a forward-simulated trajectory — unconfirmed defaults to blocked.
A red-team that measures it
Our safety layer is open source: 308 automated tests — unit, mutation, fuzz, black-box — plus live ISO/TS 15066 hazard trials. Verify it yourself on GitHub.