▧ episode_0142.mp448 MB
{ } annotations.json12 KB
◫ robot_state.mcap8 MB
Pianite
Bring every episode into context.
Connect robot recordings, first-person demonstrations, task instructions, and recorded states and actions in their original formats.
Data quality for physical AI ›
Confidence in every frame.
Audit robot demonstrations, verify task outcomes, and catch data issues at scale.
Our human reviewers handle the edge cases. Your team stays focused on building.
ROBOT DEMONSTRATION EPISODE 0142
▶Reach & grasp
00:12
P
Our experts take it from here.
✓Pianite reviewers handle the edge cases.
Pianite · Data workspaceInteractive preview
Datasets / Kitchen manipulation⋯
DATASET OVERVIEW
Kitchen manipulation
Every episode. Every camera. One source of truth.
Episodes analyzed2,048Across 3 camera streams
Quality score94.2/100↗ Ready for review
Flagged episodes118Recording & task issues
EpisodeTaskDurationStatusScore
Sample data for illustrationAll checks up to date ●
Pianite found an issue now
Instruction mismatchThe recording shows a placement, but the label describes a grasp.
Needs reviewEpisode 0142 ↗
$ pianite audit kitchen-v2
✓ Camera streams synchronized
✓ Metadata integrity checked
✓ Semantic review complete
2,048 episodes · 118 flagged
→ Report ready▌
✓ Camera streams synchronized
✓ Metadata integrity checked
✓ Semantic review complete
2,048 episodes · 118 flagged
→ Report ready▌
▧ Video
⌘ LeRobot
◫ MCAP
{ } JSON
▱ Archives
PlatformKnow what your robots will learn from.
Know what your robots will learn from.
Catch recording faults. Validate demonstrations.
Our experts resolve the uncertain episodes.
01 / Recording integrity
Catch problems in the recording.
Inspect camera streams alongside recorded states and actions. Surface corrupted frames, timing drift, and missing data with the exact evidence to review.
Every stream, in context.
Review synchronized views and motion on a shared episode timeline.
Episode 0142 / Camera review3 streams in sync
CAM 01 · FRONT00:12.400
target object · 0.98
CAM 02 · WRIST
CAM 03 · SIDE
00:12 / 00:32Action timeline
Reach
Grasp
Move
Place
0:000:080:160:240:32
⚑Instruction mismatch detected at 00:12Review
From data to clarityFrom raw demonstrations to reviewed datasets.
From raw demonstrations to reviewed datasets.
Automated checks. Targeted human review. Clear dataset decisions.
01
✓ Recording integrity Passed
✓ Camera synchronization Passed
⚑ Instruction consistency 2 issues
✓ Task completion Passed
Test the data behind the behavior.
Combine rule-based checks with AI review of labels and task completion. Surface specific failures with frames and timestamps our reviewers can inspect.
Dataset quality report ↗
94.2 Quality score
✓ Reviewed↓ Export JSONL
Our reviewers close the loop.
When an episode needs human judgment, Pianite brings in our reviewers. We handle the review and annotation corrections, so your team can keep building.
Preference data & semantic curationData that reflects your preferences.
Data that reflects your preferences.
Curated for your models. Or your next customer.
You define what good looks like.
We build the dataset.
Tell us which tasks, behaviors, and outcomes matter to you. Pianite curates semantically rich datasets around those criteria, with our human reviewers capturing the preferences that make a demonstration valuable.
Use the collection to train your models, or offer a focused dataset for sale. We handle selection, review, and annotation, so your team can focus on what comes next.
YOUR DATASET BRIEF
TaskPlace objects on a shelf
ContextEveryday kitchen environments
OutcomeSuccessful, complete demonstrations
Human preferences
Gentle contactPrecise placementMinimal retries
✦
A collection built around your intent.
Pianite-reviewedSemantic annotations · Preference labels · Quality reports
Illustrative dataset brief
01
Capture human preferences.
Our reviewers identify the demonstrations that match how you want a task performed, beyond whether it succeeds.
02
Curate by meaning.
Build specific collections around task intent, scene context, observed actions, and outcomes.
03
Package for training or sale.
Deliver focused datasets with annotations, preference labels, and quality context your team or buyers can inspect.
Built for contextUnderstand the episode.
Understand the episode.
See the evidence behind every flag.
Video, recorded motion, and task instructions tell the story together.
▧ Camera streams⌁ Robot actions▤ Annotations
01
Find failures that affect learning.
Spot failed grasps, incomplete demonstrations, idle footage, and task instructions that do not match the recording.
02
Keep the evidence close.
Trace each issue to the affected camera and timestamp, with recorded states and actions alongside the footage.
03
Our humans. Your peace of mind.
Our reviewers resolve unclear task outcomes and conflicting labels, with the full episode context. You get reviewed results without managing a review team.
ConnectivityYour data stack, connected.
Your data stack, connected.
Built to meet your data where it lives.
Object storage
Your buckets. Your recordings.
LeRobot datasets
Episodes, actions, and states.
MCAP recordings
Robot streams, in context.
JSON & JSONL
Structured results for your stack.
For buildersQuality reports your
Quality reports your
training pipeline can use.
Keep task outcomes, action timelines, and recording faults in structured outputs your team can review and export.
// Illustrative episode report
{
"episode_id": "0142",
"outcome": "unclear",
"review_status": "human_review",
"issues": [{
"type": "instruction_mismatch",
"timestamp_s": 12.4,
"camera": "front"
}]
}Better data. Better models.Better robot learning
Better robot learning
starts with better demonstrations.
Catch data issues before the next training run.