Data quality for physical AI ›

Your best data
for physical AI.

Pianite tests and curates your data autonomously.
Our humans handle the cases that need judgment.

Explore workspace ↗
Pianite Episode review · DemoExplore this episode ↗
EXO · EPISODE 000018
Right armApproach and align jawsBlue 2 block
00:24 / 00:44
▧ Video
⌘ LeRobot
◫ MCAP
{ } JSON
▱ Archives
Use cases

Better data for your next step.

01 / Training

Train on what matters.

Find clean demonstrations, useful recoveries, and the gaps behind model failures.

Explore your dataset ↗
02 / Preferences

Define better behavior.

Collect human choices for careful handling, speed, or the criteria you define.

Compare demonstrations ↗
03 / Data delivery

Give labs your best.

Curate focused collections for your own models, or prepare your strongest data for AI labs.

Build a collection ↗
Platform

Quality, built into every episode.

01 / Recording integrity

Catch problems in the recording.

Inspect camera streams alongside recorded states and actions. Check for corrupted or missing frames, timing drift, sped-up recordings, and swapped camera streams. Keep recording faults separate from task failures, with evidence for each finding.

✦
Every stream, in context.

Review synchronized views and motion on a shared episode timeline.

Dataset intelligence

Find your best data.
See what is missing.

Search by meaning.
Curate with evidence.

Find demonstrations by task, scene, action, or failure. Inspect matching frames, compare similar episodes, and build a focused selection.

01Spot repeated demonstrations.

Reduce duplication and inspect unusual behavior.

02See what is missing.

Compare coverage across tasks, conditions, and outcomes. Track collection trends by operator, with task difficulty in context.

03Turn failures into the next brief.

Use the gaps behind model failures to guide collection.

From data to clarity

From raw episodes
to usable data.

01

Connect your data.

Bring recordings, instructions, actions, and states in their original formats.

02

Run quality in the background.

Clear cases move forward. Our reviewers handle only the decisions that need judgment.

03

Deliver a focused dataset.

Keep selected episodes, corrections, preference labels, and the evidence behind each choice.

Humans, only where needed

Autonomous first.
Our humans when needed.

Our people.
Your peace of mind.

Clear findings move forward automatically. Our reviewers handle uncertain evidence against your criteria. Targeted sample review looks for missed problems.

Our team manages the review and returns each decision with its evidence. You do not need to staff a review team.

Explore the review workflow ↗
From findings to usable data

Quality you can measure.

01 / Your criteria

Define what belongs.

Set the task, desired behavior, and acceptance rules. Keep recording integrity, instruction accuracy, and execution quality as separate decisions.

02 / Evidence and judgment

Test the labels, too.

Agree on a human-reviewed reference sample for your pilot. Assess false alarms, missed issues, timestamp accuracy, and agreement between reviewers. More labels alone do not establish accuracy.

03 / Your next training run

Keep the useful parts.

Preserve recoveries and valuable segments from imperfect attempts. Deliver the selected data, corrections, and reasons against the criteria your team accepts.

Start with a scoped pilot to measure review time and accepted corrections. Test model improvements with a controlled training comparison.

Freeform preference data

Human preferences.
In your own words.

The same pair.
Different reasons to prefer each.

Define preference dimensions in natural language, such as careful handling, speed, or placement quality. Our reviewers compare full demonstrations along each dimension.

Receive each pair, preference description, and human choice in a structured dataset. Automation prepares the comparisons. Our reviewers supply the judgments for reward learning and behavior alignment.

Compare demonstrations ↗
01

Describe the preference.

Start with your criteria. Our reviewers can add task-relevant dimensions as the review evolves.

02

Compare each dimension.

Keep carefulness, speed, and result quality as separate human judgments for the same pair.

03

Train with richer feedback.

Use the labels to train rewards that account for the named preference and support controllable behavior.

Curated datasets

Give labs your best data.

You define what good looks like.
We build the dataset.

Define the tasks, scenes, actions, and outcomes you need. Pianite selects and annotates demonstrations around that brief autonomously, with targeted human review and preference labels where they matter.

Give AI labs your strongest demonstrations, or curate the best data for your own models. We handle selection, review, and annotation around your purpose.

01

Define and approve the sample.

Set the task, scene, desired behavior, and intended use. Align on sample demonstrations before the full collection.

02

Curate and review the collection.

Pianite selects and tests demonstrations against the agreed brief. Our reviewers handle uncertain cases. Each quality decision stays tied to evidence.

03

Package for training or sale.

Deliver your best data to AI labs, or keep a focused collection for your own work. Include annotations and quality evidence that explain each selection.

Connectivity

Part of your data workflow.

▱Object storage

Review new batches from your storage.

⌘LeRobot datasets

Episodes, actions, states, and versions.

◫MCAP recordings

Robot streams, in context.

{ }JSON & JSONL

Reports, annotations, and review decisions.

For builders

Clear findings.
Structured results.

Run quality review as new batches arrive. Keep recording tests, task outcomes, reviewer decisions, and dataset versions in structured outputs for your pipeline.

Preview connections ↗
// Illustrative episode report
{
  "episode_id": "0142",
  "recording_quality": "passed",
  "task_outcome": "complete",
  "review_status": "human_review",
  "issues": [{
    "type": "instruction_mismatch",
    "timestamp_s": 12.4,
    "camera": "front"
  }]
}
A closer look

Answers, before you start.

Why use Pianite if annotation tools are open source?

Annotation tools generate findings. Pianite’s service brings those findings into a managed workflow: your criteria, targeted human review, corrections, and a curated delivery for your next training run.

How will we assess quality?

Agree on a human-reviewed reference sample and acceptance criteria for the pilot. Measure false alarms, missed issues, timestamp accuracy, and reviewer agreement. Any claim about model performance needs a controlled training comparison.

Will you discard every failed demonstration?

No. Recording corruption, label errors, and task failures are different problems. Useful recoveries and valid segments can remain in a collection when they fit your criteria.

What does Pianite evaluate?

Recording integrity, annotation consistency, task outcomes, and demonstration behavior. We connect each finding to the relevant frame, timestamp, and task criteria.

How much work needs human review?

Automation is the default. Pianite escalates uncertain findings and decisions that need human judgment. The review scope follows your quality criteria, without a fixed claim about the percentage of automated decisions.

Who handles the human review?

Pianite runs most quality work autonomously. Only cases that need human judgment reach our reviewers. We manage their work, corrections, and documented decisions. You do not need to staff a review team.

What is preference data?

Freeform preference data captures a human choice for each named dimension of behavior. A reviewer can prefer one demonstration for carefulness and another for speed. Pianite manages the comparisons and preserves those separate labels.

Can you curate a dataset for a specific use case?

Yes. Start with the tasks, scenes, actions, outcomes, and preferences you need. We align on sample demonstrations, then curate and review the collection against that brief.

Can I prepare a dataset for sale?

Yes. We curate a focused collection with annotations, preference labels, and quality reports. The project brief also defines source documentation, usage terms, and delivery requirements.

How does this fit into a training workflow?

Use quality findings to select episodes, use preference labels to guide behavior, and use coverage gaps to plan collection. Versioned outputs keep the next training dataset tied to its review history.

Better data. Better models.

Better data.
Without the busywork.

Autonomous quality. Our humans for the hard calls.
Give labs your best data, or curate it for your own goals.

Explore the workspace
Let’s build better data

Meet Pianite.

Tell us about your data, quality criteria, human preferences, or dataset brief.

This opens your email app with a draft. Your message is sent only when you send it.

Enlarged Pianite Studio screenshot

Captured from the working Pianite public dataset workspace.