Gap read
6 ways models fail in the real world.
Physical AI models rarely fail at random. They fail in six repeatable ways, and each one has a different fix. More of the same does not help.
The six failure types
Name the failure. Then fix it.
01
Gaps
The model has never seen it.
- Where it shows up
- A home robot meets a kitchen layout, appliance or object it was never shown.
- How Ancor closes it
- We capture the missing scenes in real homes, on demand.
02
Drift
The model knew it once. Then the world moved on.
- Where it shows up
- A delivery or driving model trained last year meets new streets, seasons and signage.
- How Ancor closes it
- We recapture the same places and tasks as they look now.
03
Thin coverage
The model knows the category, but not in enough depth.
- Where it shows up
- A picking model knows boxes, but has seen only a handful that are crushed, wet or shrink-wrapped.
- How Ancor closes it
- We capture many real variations of the exact case that fails.
04
Conflicts
Sources disagree, and the model learns the average of two truths.
- Where it shows up
- A world model sees objects fall, slide or deform differently across sources.
- How Ancor closes it
- We capture clean ground truth with camera pose, metric depth and full provenance.
05
Wrong data
The model learned from inputs that were simply wrong.
- Where it shows up
- Mislabeled actions, misaligned frames, or synthetic scenes that break real physics.
- How Ancor closes it
- We replace them with verified real-world observations.
06
Narrow view
The model has seen the world from too few places, people and cultures.
- Where it shows up
- A video or human model works in San Francisco and fails in Lagos, Seoul or a rural village.
- How Ancor closes it
- We capture across our contributor network in 174 countries.
A single failure often mixes more than one type. Narrow view will matter more as regulators ask how representative a model’s training really is.
The gap read
One failure case. One week.
Send us one case where your model breaks. Within a week we come back with what type of failure it is and how we would close it.
- 01You send one failure caseA clip, a scene description or an eval result. Whatever shows where the model breaks.
- 02We classify itWhich of the six failure types it is, and how much of each.
- 03We scope the fixThe real-world scenes to capture, where, and how many.
- 04We show a sampleReal observations of your failure case, with camera pose, metric depth and provenance.
- 05You decideRun a capture pilot, or keep the read and walk away. No cost, no commitment.
Start here
Send us one failure case.
We are working with a small number of teams this quarter, so each one gets capture built around its model. We reply within 24 hours.