One layer between an AI brain and any robot body.
nex-ON perceives the real scene, picks the right tool, and acts — directed entirely in plain human language. Below is everything running today, and the architecture that lets it drive any robot next.
Robots are powerful.
Deploying them is brutal.
Putting a robot to work today means one of two slow, expensive paths — and neither adapts when the real world doesn't match the plan.
The Teach Pendant
A specialist hand-jogs the machine, recording waypoints one by one. Slow, manual, and only as good as the person holding it.
Offline CAD Programming
An engineer writes a program for every new part or task. Weeks of work that breaks the moment reality deviates from the model.
Scarce Skilled Labor
Both paths depend on specialist expertise that is hard to hire and harder to keep. The people who can deploy a robot are rarer than the robots.
High-Mix, Low-Volume Work
Short runs and constant changeovers can rarely justify the cost of a robotics integrator. The robot sits idle; the work goes manual.
Unstructured Environments
Rigid programs trust a model file. When the real part sits a few millimeters off, the robot fails — which is exactly what blocks humanoids and mobile robots from the real world.
So stop programming the robot.
nex-ONreplaces “program the robot” with “talk to the robot.” It sees the actual scene in front of it — not a model file.
Most robotics companies
build one robot for one job.
nex-ON inverts that. It sits between an AI brain and a robot body — the layer that turns “understand the goal” into “perceive the world, pick the right tool, and act.”
Because perception, tooling, and control are interfaces rather than fixed implementations, adding a new robot or a new skill means registering a new tool — not rebuilding the system.
Beachhead first, platform underneath
We prove the platform on the hardest, most valuable near-term task we could find — one that demands everything at once: sub-millimeter perception, safe real-world actuation, and non-expert operability. Nail that, and the same platform generalizes outward.
See the first deploymentThe Brain
A large language model reasons about the goal and orchestrates the work — deciding when to look, what to measure, and where to act.
The Capabilities
Modular, swappable tools the brain can call: vision models, measurement, motion, manipulation, sensor feeds. Register a new tool, gain a new skill.
The Body
Any robot. Today it drives a collaborative arm. The same platform is architected to drive a humanoid, an AMR, or a fleet of mixed robots.
In principle, any robot plus the right tools equals any task — all directed through natural conversation.
A conversation,
not a program.
The operator speaks. The model decides when to look, what to measure, where to act, and how to move — then does it and reports back out loud.
One loop, six moves
Voice or text, hands-free
You speak the goal
Push to talk, in your own language. The session can be locked to one language so background chatter can't hijack it. An optional barge-in mode lets you interrupt the robot just by starting to talk.
Agentic tool-calling loop
nex-ON reasons and chooses
The model decides, mid-conversation, when to look through the camera, what to measure, when to move, and when to act. It composes the tools it needs rather than following a fixed script.
Open-vocabulary vision & measurement
It perceives the real scene
Ask for any object in plain words — “metal tube,” “flange,” “plate” — with no per-class training. nex-ON fuses the image with aligned depth to estimate real length, width, and distance in millimeters.
Safety by construction
It checks before it moves
A dry-run reachability check reports whether a move is feasible before the arm moves. Dangerous actions must be deliberately armed. Speeds default low, and per-axis motion locks constrain what can change.
Perception mapped to motion
It acts
A calibrated hand-eye transform turns “the pixel I see” into “the exact 3D point to move to.” The robot executes against what it actually observed — not a model file.
Spoken back, immediately
It tells you what happened
Replies start speaking after the first sentence while later sentences are still being generated, so it stays responsive even on long answers. The terminal shows only the conversation; diagnostics go to logs.
Everything below
is running today
Conversational Orchestration
nex-ON runs an agentic tool-calling loop — choosing mid-conversation when to look, move, and act, then narrating the result.
Voice In and Out, Multilingual
Push-to-talk recognition and streamed speech. Lock a session to a single language so background chatter can't hijack it. Optional barge-in to interrupt.
Open-Vocabulary Vision
Ask for any object in plain words — no per-class training. The detection backend is a swappable interface.
Real-World Measurement
Fuses image, aligned depth, and camera intrinsics to estimate a part's length, width, and distance in millimeters.
Color-Guided Pathing
Detects red markers, dots, and drawn or taped lines by color, then moves to or traces them — including shortest-path multi-target routes.
Motion Safety Throughout
Separate working and positioning speeds, per-axis motion locks, and a dry-run reachability check before the arm moves.
Hand-Eye Calibration
A calibrated camera-to-robot transform turns the pixel it sees into the exact 3D point to move to.
Sensor Integration
Depth and image streams register as tools the reasoning loop can call. New sensors join the same way.
Operator-Friendly Runtime
The terminal shows only the clean conversation. All diagnostics stream to timestamped logs.
Interfaces, not fixed
implementations
The brain calls modular tools. The tools drive any body. Swap a detector, add an end-effector, register a robot — the platform compounds instead of being rebuilt.
Human goal
Plain language
nex-ON
The reasoning brain
calls modular tools
Vision
Detection
Sensors
Integration
Motion
IK & control
Tooling
End-effector
New tool
Extensible
any body
Proven today·Architected, on the roadmap
One platform.
Any body.
We are precise about what ships today and what the architecture is built for. The modularity is real in the codebase. The additional robot bodies are the roadmap.
Tasks
Welding
OmnicronAssembly
Inspection
Material handling
Mobile manipulation
Proven today
The full stack running on a real collaborative arm — voice orchestration, open-vocabulary vision, real-world measurement, safe dry- and live-gated actuation, color-guided pathing, and hand-eye calibration — proven end to end as Omnicron.
Architected for, not yet shipped
The same orchestration and tool interfaces extending to humanoids, AMRs, and additional end-effectors and sensors. The swappable detector, tool-based capabilities, and abstracted motion exist today. The additional bodies do not — yet.
Where we sit
vs. task-specific robotics
Single-process systems and no-code programming tools solve one job on one class of machine.
Our edge: nex-ON isn't tied to one process or to one kind of arm. It's horizontal. Our first application is one deployment, not the ceiling.
vs. embodied-AI & general robotics
Humanoid and robot-foundation-model companies build robot bodies or end-to-end learned control policies — capital-intensive, hardware-heavy, often single-embodiment.
Our edge: We are not a robot manufacturer and not a monolithic learned policy. We sit above any body and compose proven perception and motion tools under an LLM's reasoning.
Works with robots that already exist, from many vendors.
Interpretable and controllable — discrete tool calls with safety gates and dry runs, not an opaque neural policy.
Swappable by design — a better detector or a new sensor drops in without a rewrite.
Everyone else is building the robot's body or its reflexes.
nex-ON is the deployment platform that lets any body do any job.
What actually changes
on the floor.
Steps disappear from the deployment path.
6steps to deploy
3steps to deploy
No programming step
No specialist required
Adapts to the real scene
Rehearse before you commit
Low integration cost
Every tool compounds
Start with a single cell. The tools you register there are the same tools the next robot calls.
The questions
buyers actually ask.
Straight answers on hardware, setup, safety, and where the AI runs. Something not covered here? Ask us directly.
nex-ON is robot-agnostic by design. Today it runs on a FAIRINO FR5 collaborative arm — that pairing is Omnicron, our welding application. Orio, the mobile application, is built on NVIDIA edge AI hardware (Jetson Thor / Orin and DGX Spark). Because perception, tooling, and control are interfaces rather than fixed implementations, the same stack is architected to drop onto other arms and bodies.
Stop programming.
Start talking.
See nex-ON find a part, measure it, and rehearse a full pass — directed entirely by voice. Then talk to us about the robot you want it to drive next.
Bring Your Own Robot
We're not a robot manufacturer. nex-ON is built to sit above hardware you already own.
Rehearse First
Every dangerous action runs dry by default. Nothing is energized until you arm it.
Engineering-Led Pilots
Start on a single cell. Our team is in the loop from calibration to first live pass.