The public demonstration and research layer for Physical AI evaluation. Each lab is useful to one engineering or operational decision, runs without an account, and says plainly what is live, what is synthetic, and what still needs a real site.
Define what has to be tested before any robot runs, and see the Evidence Record that test would produce.
Each opens the console in the state that answers its question. Filter by decision stage, or by whether the data is real.
LAB-01 · PlanLive demoQ. Which few sites test the most conditions my robot will face?
The test site planner picks K sites by condition coverage inside an operating envelope, against the K most suitable sites. On the demo, 5 planned sites cover 20 of 28 conditions (71%); the top 5 by suitability cover 9 (32%).
A validation budget buys coverage, not repetition.
LAB-02 · AssessLive demoQ. What grades does a ground robot meet on this route, and where?
Elevation and grade along a 7.15 km route from the Mount Bonnell ridge to the river at downtown Austin: 241 m to 144 m, steepest −20.7%, 72% of the route under 5% grade.
Traversability is a route property, not a site property.
LAB-03 · AssessLive demoQ. How do the weights and raw conditions turn into a suitability score?
The score distribution under the current weights, how one criterion's raw values are transformed into suitability, and a query that highlights the parcels meeting a stated condition (slope under 8% and suitability over 60: 46 of 432).
A score is only defensible with its transformation and weights attached.
LAB-04 · DiscoverLive demoQ. What do the best candidate sites have that the rest do not?
The top N% of sites against the rest: the ranked score curve and cut-off, and a criterion profile showing which conditions drive the difference, beside a map of the classification.
A shortlist should come with the reason it is a shortlist.
LAB-05 · DiscoverLive demoQ. Where are the ridges, valleys and drainage on a candidate site?
Colour relief in 3D, stretched over the elevations in view, anywhere on Earth: blue valleys to red ridges, with hillshade.
Terrain screening needs no local data and no site visit.
LAB-06 · RecordLive demoQ. What does the evidence layer look like once a site is instrumented?
The registered substrate channels on one site: traffic density, a logged run with GNSS error ellipses that widen in a degraded stretch, IMU acceleration ticks, and radar returns coloured by Doppler velocity.
This is the shape of the data an Evidence Record is built from.
These answer questions only measured evidence can answer. They will appear here as validation engagements produce them, and not before.
How does an AMR's stopping distance change on a wet surface in low light?
Needs · A robot, a yard section, and surveyed ground truth
Where does localisation degrade on a real site, and by how much?
Needs · An instrumented run with an RTK reference
Where does a simulated run disagree with the same run in the field?
Needs · A customer's simulation scenario and a matching field run
Each phase is built when the one before it is useful. Statuses use the same four words as the rest of the site.
Show credible workflows and current software capability, without a physical lab.
Turn demonstrations into repeatable technical workflows.
Expand only when demand, equipment, sites and measurement capability justify it.
Hashes and version numbers alone do not make evidence trustworthy. What matters is keeping the link between source data, processing steps, configuration and reported result, which is what the Evidence Record is designed to carry.
Field notes and technical briefs. Established results, hypotheses and proposed methods are kept apart.