AI-NATIVE ROBOTICS · PEOPLE AND ROBOTS, SIDE BY SIDE
Day one, we scan the building. Day two, the AI-native robot works a supervised shift.
RoboCoworker turns a building's layout, objects, procedures and safety rules into a tested Site Pack, so a general-purpose humanoid or ROS-capable robot becomes a safe co-worker in your specific site.
One site pack, versioned together
- 01Spatial graphOpenUSD
- 02Object catalogueUSD assets
- 03Task graphSkill sequences
- 04Safety envelopePolicy
- 05Acceptance suiteIsaac Lab
THE PROBLEM
Robots are general. Sites are not.
Foundation models give robots general skills. Nothing gives them your building.
11 months.
Figure 02's pilot at BMW Spartanburg, judged on cycle time, placement accuracy and interventions.
2+ weeks
Siemens’ estimate of how long training a humanoid for a factory can take.
~1 year
What even basic automation projects can need, with heavy on-site testing, to reach production reliability.
Sources: Figure AI · CDOTrends (Siemens) · The Robot Report
The insight
Turn site knowledge into data robots can train on.
Today, what a robot learns about a building lives in an integrator's head and a pile of config files. Make it versioned and testable, and it becomes portable, auditable and cheaper each time it is reused.
PORTABLE
Moves between sites and between robots.
TESTABLE
Rehearsed in simulation before the robot arrives.
CERTIFIABLE
A safety record you can show a reviewer.
THE PRODUCT
The Site Pack
One bundle that makes a general robot useful in one specific building. We don't build the robot or train the foundation model. We make your preferred robot useful in your building.
01
Spatial graph
Zones, doors, thresholds, no-go areas and floor friction classes.
Gaussian-splat scan plus LiDAR collision mesh
02
Object catalogue
What this site handles: shape, mass, grasp points, fragility.
Measured, simulation-ready assets
03
Task graph
The procedures staff already follow, compiled into skill sequences.
Staff interviews plus teleoperated demos in the twin
04
Safety envelope
Speed by zone, separation distances, stop and escalation rules.
Declared policy, stress-tested in simulation
05
Acceptance suite
Site-specific tests the robot must pass before it enters.
Simulation scenarios with pass thresholds
HOW IT WORKS
Scan to shift in 48 hours, trained on cloud GPUs overnight.
Two days on site; the heavy work runs overnight on cloud GPUs. Clock times are design targets that the proof of concept measures.
Scan
Walk the site once with LiDAR and RGB. Record procedures with staff. Mark no-go zones.
Twin
Reconstruct as Gaussian splats plus a collision mesh. Place objects from the catalogue.
Rehearse
Start from the nearest site archetype. Run thousands of randomised episodes overnight.
Certify
Pass the acceptance suite and envelope gates. Issue a signed record.
Shift
Shadow mode first, then supervised autonomy beside staff. Telemetry flows back.
Every intervention on shift becomes a new scenario in the acceptance suite. The pack gets better at this site, and the archetype gets better everywhere.
USE CASES
Jobs a humanoid can do now.
Factory: line-side replenishment
A mixed assembly cell where people and the robot share one aisle.
The job
Take a parts tote from the supermarket rack, cross a 15–25 m pedestrian aisle, and swap it into the line-side lane.
On the real robot
Gantry, then tether, then free. One shadow shift beside a person doing the same job.
| KPI | Pass | Stretch |
|---|---|---|
| Task success per attempt | ≥ 90% | ≥ 97% |
| Cycle time vs human baseline | ≤ 2.0× | ≤ 1.3× |
| Interventions per operating hour | ≤ 2 | ≤ 0.5 |
| Line stoppages caused by the robot | 0 | 0 |
| Scan start to first useful task | ≤ 5 days | ≤ 48 h |
Thresholds are proof-of-concept proposals, set with each partner.
Hotel: back-of-house runs
Staff corridors first; guest corridors only after the safety envelope holds.
The job
Collect an amenity caddy from the linen room, deliver it to a staged room door, and bring a bagged linen sack back.
On the real robot
We pick jobs this body can do now: not beds, since a king duvet is wider than any humanoid’s span.
| KPI | Pass | Stretch |
|---|---|---|
| Right item to the right door | ≥ 90% | ≥ 98% |
| Separation-distance breaches | 0 | 0 |
| Yields when a person approaches | ≥ 95% | ≥ 99% |
| Attendant walking saved per shift | ≥ 30 min | ≥ 60 min |
| Re-certify after furniture moves | ≤ 4 h | ≤ 1 h |
Thresholds are proof-of-concept proposals, set with each partner.
Construction: fit-out patrol and staging
Indoor fit-out only: the H1-2 is a research platform, not a site-hardened machine.
The job
Walk a 60–100 m patrol route capturing images and LiDAR at waypoints; carry a fixings box or tool bag to a marked zone.
On the real robot
Harness or gantry on every real run; floor closed to other trades during trials; a spotter holds the e-stop.
| KPI | Pass | Stretch |
|---|---|---|
| Patrol completed without assist | ≥ 90% | ≥ 98% |
| Arrested falls per 10 km walked | ≤ 1 | 0 |
| Waypoints captured | ≥ 95% | ≥ 99% |
| Morning re-scan to re-certified pack | ≤ 3 h | ≤ 1 h |
| Materials staged to the correct zone | ≥ 90% | ≥ 97% |
Thresholds are proof-of-concept proposals, set with each partner.
Positioning
The site layer between foundation models and real buildings.
RoboCoworker stays neutral across the layers beneath it. When a better model or robot arrives, the site record carries over.
-
Site operator
Factory cell · hotel back-of-house · fit-out floor. Needs a working shift, not a capability.
-
RoboCoworker
Site Pack · acceptance suite · safety envelope · re-certification when the building changes
-
Foundation models
Skild Brain · π0.5 · GR00T N1.6 · Field Foundation Models
-
Simulation and data
Isaac Sim and Lab · Omniverse NuRec · Lightwheel assets · Parallax twins
-
Robot hardware
Unitree H1-2 and H2 · Agility Digit · Figure · Apptronik
PARTNER WITH US
Run a joint proof of concept.
We're looking for labs, site operators and robot makers to validate the Site Pack in mock factory, hotel and construction sites first, then on partner sites.
