The Platform
NEXUM
The platform layer for physical AI. One capture standard, curated data from many people and places, models built to transfer across machines, and an on-demand skills marketplace — open to any hardware built to it.
The Problem
Machines are getting better.
Data isn’t keeping up.
Capable robot hardware is getting easier to build. What doesn’t get easier is the data: years of demonstrations, collected one robot at a time, mostly inside a handful of labs.
That leaves smaller teams with good machines that don’t know how to do much, and leaves the best models trained on narrow slices of the world. NEXUM is built to fix both — by collecting from people instead of robots, and by sharing what is learned across every machine on the platform.
Architecture
Five layers. One platform.
01
Capture standard
One schema for egocentric video, hand pose, touch and motion. Our capture kit records to it natively — and so can capture hardware from partners. Every stream lands in the same shape, whoever built the rig.
{
"source": "VISUS + TACTUS + GRESSUS",
"streams": ["ego_video", "hand_pose", "touch", "imu"],
"setting": "workshop",
"task": "plane a board"
}02
INDEX
The curation layer. Every stream is time-aligned, labelled, quality-scored and tracked by where it came from — the person, the place, the task, the hardware. Diversity is measured, not assumed.
✓ ALIGNED
✓ LABELLED
✓ SCORED
✓ TRACED
03
Foundation models
Models trained across many people, places and machines, built to transfer to any embodiment instead of overfitting to one. The more varied the data, the more general the model.
04
Skills marketplace
Capabilities packaged as on-demand skills and utility packages. Any machine subscribed to the platform can call them, so new hardware launches with a library on day one.
machine.call("fold_garment") // any subscribed machine machine.call("open_door") machine.call("sort_items")
05
Orchestration
When a job needs more than one machine, NEXUM breaks it into subtasks and coordinates the fleet — whoever built each machine.
map() → fetch() → place() → verify()
How It’s Shared
Built to be built on.
AVAILABLE TO PARTNERS
- — The capture schema, so any rig can record data that joins the platform
- — The hardware integration spec: capability declaration, telemetry and instruction interface
- — Access to skills for every machine built to the spec
RUN BY AI MACHINAE
- — INDEX, the curation layer behind the shared dataset
- — NEXUM foundation models
- — The skills marketplace and fleet orchestration
The Flywheel
Every machine makes
every other machine smarter.
More contributors
↓
More diverse data
More diverse data
↓
Better models
Better models
↓
More useful machines
More machines
↓
More contributors