Deformable Standards
A written standard for each task: how it should be done step by step, what counts as a defect, how to score the result. Every operator, labeler and model judges the same way.
DFRMBL · SPEC 01 · REV A
Expert judgment for physical AI
Robots can fold. We tell them when they got it wrong.
Clothes, towels, bedding, packaging, food — for soft things there’s no rule for “done right.” Deformable turns the judgment of people who’ve handled these materials for years into task standards, expert labels and a judge model that catches the mistake while it happens.
Flat 01 · T-shirt · front
Review
- 1grasp here ✕
- 2should grasp ✓
- 3collar off-center
✕ GRASP_OFFSET_FROM_SEAM · ~5 cm → collar off-center
VERDICT: REWORK · grasp at equal distance from both side seams
Key figuresSpec 01
- 0 / 21
- causes of failure a frontier AI model named in our blind test
- 17
- failure types in our folding taxonomy
- 10+ yrs
- of production expertise behind every label
- 5
- material categories: garments, towels, bedding, packaging, food
01 / Problem
More data doesn’t fix a bad fold. Judgment does.
Robots learn physical tasks from thousands of recorded attempts. Someone has to decide which attempts were done right and why the rest failed. For a box, that’s easy — it’s in the bin or it isn’t. For a towel, a duvet or a T-shirt, there’s no formula. Teams check the last frame by eye, or use AI judges that are close to guessing on anything involving contact. And the mistakes that matter happen a minute before they show.
Flat 01 · T-shirt · frontDefect
- 1grasp ~5 cm off the side seam ✕
- 2grasp on the seam ✓
- 3collar off-center at the end
02 / Product
Three layers of judgment. One loop.
Deformable Labels
Expert labels on robot video: the moment it went wrong, the cause in material terms, the fix, marked on the frame. Ready for filtering demos, training reward models and fine-tuning.
Deformable Judge Early access
A vision-language model trained on expert labels. It watches a robot work, flags the mistake, explains why and shows where to grasp instead — so a human doesn’t have to babysit the robot.
03 / Services
From your first failed rollout to a judge in your training loop.
- 01
Failure review Start here · Free
Send 20 failed rollouts. Within about a week, get a report: the moment each attempt went wrong, the cause, the fix, marked on the frame — and the 3–5 reasons your robot keeps failing.
- 02
Expert labeling
Production experts label your robot data at volume: every defect and correct action, why it happened, what to do instead. Structured JSON aligned to your episodes and timestamps.
- 03
Task standards & taxonomy
A written standard and failure taxonomy for your tasks, built with production experts. Brief operators, align labelers, score models.
- 04
Policy evaluation
Before you ship a new policy, run its rollouts past expert judgment. Success rate by failure type, not just pass/fail.
- 05
Judge API Early access
Deformable Judge on your rollouts: per-episode verdict, failure type, cause and fix — as a score for filtering data or a reward for RL.
- 06
Expert data collection On request
Demonstrations recorded by production experts on real factory floors in Central Asia, with full material specs labs never capture.
Pricing per project. Start with a free failure review.
04 / Sample review
Same fold. Two judges.
Flat 01 · T-shirt · frontTechnologist
Technologist
- 0:16
- Holds the collar, not the shoulder-sleeve seam junction → diagonal first fold.
- 0:36
- Grasp ~5 cm off the side seam → edge folds under.
- 1:20
- Grasp ~10 cm off the side seam → asymmetric fold.
- 1:42
- Collar off-center — caused by 0:36 and 1:20.
- Verdict
- Rework. Fix: grasp at equal distance from both side seams.
Frontier AI model
- 0:36
- Third shake — looks fine.
- 1:20
- Folds the strip in one move.
- 1:42
- Small flap near the neckline.
- Verdict
- Acceptable
Based on a real episode from an open robot-folding dataset and a blind test of a frontier AI model.
05 / Research
We publish what generic judges miss.
Blind test: frontier AI vs. a garment technologist
Two “high-quality” episodes from an open robot-folding dataset. The technologist found 21 defects and named every cause in garment terms. A frontier AI model named none, missed that the collar slid because of grasps a minute earlier, and called both folds acceptable.
| Measure | Technologist | Frontier AI model |
|---|---|---|
| Causes named in garment terms | 21 | 0 |
| Linked the off-center collar to earlier grasps | Yes | No |
| Verdict | Reject both | “Acceptable” |
