Skip to main content
CUPEL
CUPEL
FIND YOUR WAY
Request a pilot

Find the footage worth training on.

PLANNEDGate A: refined vs. random clips on a real SO-101 arm.

Know what your footage is worth.

Cupel scores every clip and keeps the ones worth training on.
We sell no footage or models, so we have no reason to grade yours up.

Refine 400 hours of kitchen footage from our bucket. Keep the clips worth training on.

Choose a refinery job
Choose a job. Point Cupel at a bucket.
SIMULATED DEMO · THE PIPELINE IS PLANNED

Refinery

Nine plain, rerunnable stages from raw video to a training-ready set.

EXPLORE

Coverage report

See which tasks, objects and settings a batch covers, and where it runs thin.

EXPLORE

Real-robot proof

One policy on refined clips, one on random. Tested on an SO-101 arm.

EXPLORE
Your bucketFootage never leaves it
Every clipCarries a consent record
Open formatsLeRobot, WebDataset, parquet
One ledgerEvery batch tied to an outcome

BUILT ON OPEN TOOLS

ffmpegFAISSLeRobotMediaPipe
#01 / THE REFINERY

From raw footage
to data you can train on.

refine.sh
PLANNED INTERFACE · NOT BUILT YET
READ THE ARCHITECTURE
ONE BATCH. EVERY DECISION ON THE RECORD.

Raw footage in.
A proven training set out.

The bucket becomes a manifest.
The manifest becomes a dataset.

Point Cupel at your bucket. Every clip is cut to 2–4 seconds and carries its consent record ID from the first step.

INPUT / manifest.parquet
#02 / WHAT YOU GET BACK

Every batch comes back with its evidence.

kitchen-batch / output
BATCH OUTPUT
export
labels
reports
eval
manifest.yamlsaved
Sample output · illustrative valuesYOUR BUCKET
HOW WE WILL MEASURE IT
WHO IT’S FOR

One assay for both
sides of the trade.

Labs buying footage and teams collecting it.
The same score, the same published record.

01

Robot labs

Know what a vendor’s footage is worth before you buy it.

  • Audit a vendor sample against your task list.
  • Find coverage gaps before they cost a training run.
EXAMPLE OUTPUTvendor-audit.mdRead the plan
02

Data collectors

Show buyers your footage is worth training on.

  • Refine raw capture into a clean, consented set.
  • List it with a track record, not a pitch deck.
EXAMPLE OUTPUTlisting-report.mdRead the plan
03

Research teams

Spend GPU-hours on the clips that teach.

  • Drop near-duplicates and idle footage early.
  • Compare a learned score against simple filters.
EXAMPLE OUTPUTab-result.mdRead the plan
#03 / PROOF, NOT PROMISES

A score is only worth what it predicts.

Real-robot A/B/Cgate_a
01 / THE SETUPThree arms, equal hours

Refined footage, random hours and simple filters.

02 / THE RUNOne policy per arm

Trained the same way, tested on an SO-101.

03 / THE GATE10+ pp over random and ahead of filters

Not yet run. The result is published either way.

gate_a → planned
Every row records its score version.PLANNED
The sample result is not logged yet.
READ THE EVALUATION PLAN
NEUTRAL BY DESIGN

Your footage stays yours.
One assay for everyone.

Customer footage stays in the customer’s bucket. No code path sells, exports, or shares raw footage between customers.

Robot lab
HOW NEUTRALITY WORKS
OPEN TOOLS, OPEN FORMATS

Built on tools
you already run.

We grade footage. We never sell it. The assayer can’t also be the miner.
01NeutralityNo raw footage moves between customersPRINCIPLE
Every clip carries a consent record. Faces and screens are always blurred. Neither step is optional.
02ConsentChecked at ingest, kept in every rowPRINCIPLE
We publish the track record every quarter, including the batches where the score got it wrong.
03Track recordScore version recorded on every resultPRINCIPLE

Start with one batch
from your bucket.

Cupel is pre-build. Talk to us about an early pilot.

See the refinery Request a pilot