PHYSICAL AI CUSTOM DATASET STUDIO

Human demonstration data for Physical AI.

RoboWorkData designs, records, anonymizes, annotates, quality-checks, and licenses structured human task demonstrations for robotics, automation, and embodied AI teams.

Custom protocols · Multiview recording · Structured annotations · Human QA · Commercial licensing

EPISODE: RWD-SORT-000042 // PREVIEW
Interface preview
EGO_CAM_VIEW [chest]
1080P // 30 FPS
FRONT_CAM_VIEW [fixed]
1080P // 30 FPS
SIDE_CAM_VIEW [workspace]
1080P // 30 FPS
Privacy: Cleared
QA Grade: A
Active HandRIGHT // GRASP_CONTACT
Active ObjectPLASTIC_BOTTLE_04
ANNOTATION TIMELINEEPISODE RUNTIME: 8.4S
0.00S - 2.20SAPPROACH_OBJECT
success
2.20S - 3.50SGRASP_OBJECT
success
3.50S - 6.80S *MANIPULATE_SORT
active
6.80S - 8.40SRELEASE_RESET
queued
RoboWorkData Schema Dictionary v0.1LBL_LEVEL_02 Temporal Segment

Target Domains

Who RoboWorkData serves.

We support engineering groups and research labs building the next generation of physical automation.

Humanoid Robotics

Demonstrations tailored to dual-arm manipulation and whole-body locomotion pretraining.

Service Robotics

Commercial facility routines, tidying, and tool deployment sequences.

Household Robotics

Domestic manipulation tasks inside controlled mock apartment environments.

Cleaning Robotics

Surface wiping, spill management, and vacuum routines under varied angles.

Warehouse & Logistics

Package handling, parcel sorting, and shelving tasks in mock storage hubs.

Foundation Models & Embodied AI

General-purpose dataset structures designed to feed Vision-Language-Action (VLA) architectures.

World Model Research

Multiview environment and physics sequences for structural world prediction training.

Computer Vision Teams

Highly annotated temporal boundaries, object coordinates, and hand keypoint markers.

AI Data Companies

Consent-clean, ethically gathered human task demos to diversify broad data catalogs.

Universities & Research Labs

Academic-aligned custom pilots to validate imitation learning and manipulation theories.

The Data Challenge

Physical AI models cannot scale on fragmented, synthetic, or scraped video logs.

Many teams train policies on unstandardized datasets that suffer from severe limitations:

  • Legal & Ethical liabilities: Web-scraped clips lack explicitly documented, commercial-use contributor consent.
  • No annotation depth: Raw video files miss temporal action boundaries, grasp-contact triggers, and object states.
  • Environmental bias: Fixed camera angles and single-view logs fail to generalize across real-world workspaces.
The RoboWorkData Solution

Rigorous, consent-documented human demonstration workflows.

We design custom protocols that isolate physical primitives and output structured, release-ready dataset packages:

  • Multiview recording: Synchronized ego-worn (chest/head) and static environmental views capture the full workspace state.
  • 100% consent-clean: Every session is verified against a versioned digital consent agreement with paid participants.
  • Complete annotation layer: Level 1 metadata down to Level 7 outcome and recovery logs, formatted in parquet and JSONL.

Capabilities Preview

First dataset focus: Service Work.

A preliminary catalog showcasing task families collected in controlled test apartments and facility service mockups.

View Example Dataset
Facility Service Work
Privacy reviewed

Service Work Demonstration Dataset v0.1

The flagship RoboWorkData collection: consent-based human task demonstrations spanning cleaning, tidying, object handling, and light logistics in controlled environments.

CLIPS312
TASKS20
ENVS3
QA: Human-verified QA
View Dataset
Cleaning
Privacy reviewed

Cleaning & Surface Interaction Dataset

Focused demonstrations of surface wiping, spill response, and simulated floor-cleaning routines across mock service environments.

CLIPS118
TASKS8
ENVS2
QA: Human-verified QA
View Dataset
Mock Apartment
Privacy reviewed

Household Tidying Mock-Apartment Dataset

Tidying, folding, and organizing demonstrations captured in a furnished mock apartment built specifically for controlled data collection.

CLIPS96
TASKS7
ENVS1
QA: In QA
View Dataset

Operating Pipeline

From protocol design to validated delivery.

How RoboWorkData designs, collects, reviews, and packages custom demonstration datasets.

01

PHASE_01

Requirements

Customer defines target models, environments, task families, and camera perspectives.

02

PHASE_02

Task Design

Ontology design separating distinct action verbs, start states, and target configurations.

03

PHASE_03

Collection Protocol

Specification of cameras, workspace markers, lighting setups, and allowable variations.

04

PHASE_04

Contributor Selection

Filtering pool by location, device types, languages, and upload bandwidth.

05

PHASE_05

Consent & Onboarding

Contributors complete legal consent logs and pass standardized tutorial tests.

06

PHASE_06

Multiview Capture

Continuous take recording across ego-wearable and static environmental angles.

07

PHASE_07

Data Ingestion

Lossless original uploads securely stored in private cloud storage.

08

PHASE_08

Synchronization

Aligning separate camera feeds using software markers and frame-alignment matching.

09

PHASE_09

Privacy Review

Automated heuristics flag and human annotators redact face and display identifiers.

10

PHASE_10

Annotation

Applying temporal action boundaries, object labels, and manipulation events.

11

PHASE_11

Human QA

Auditing submissions against the nine-point verification guidelines.

12

PHASE_12

Dataset Packaging

Formatting metadata as CSV, JSONL, or Parquet alongside synchronized clips.

13

PHASE_13

Commercial Licensing

Releasing cleared data with absolute ownership guarantees for B2B training.

14

PHASE_14

Scale-Up

Onboarding wider geographical pools to capture environment and lighting variance.

Metadata & Quality Assurance

Structured annotations, not raw footage dumps.

Every video demonstration is run through our redaction and labeling queue, verified by human reviewers against our 9-point criteria before delivery.

  • Task verb ontology and natural language instructions
  • Precise temporal segments (start/end millisecond timestamps)
  • Workspace environment, camera specification, and active object labels
  • Privacy clearing blurs applied on screens, badges, and faces
  • Quality grades (A · B · C · Reject) with structured check logs
STUDIO_ANALYST // VERIFYING_QA

Annotation Level 2 Preview

{
  "episode_id": "RWD-SORT-000042",
  "start_time_s": 4.20,
  "end_time_s": 6.85,
  "action": "pick_up",
  "active_object": "plastic_bottle",
  "active_hand": "right",
  "outcome": "success",
  "confidence": 0.97
}
Become a RoboWorkData Contributor

Record safe workspace tasks. Get paid for accepted clips.

Help create the structured real-world task demonstrations used to develop the next generation of robotics and Physical AI systems. Selected contributors receive clear assignments, onboarding, compensation, and project-specific recording equipment.

01 // APPLY

Submit your application with gear specs.

02 // ONBOARD

Sign consent docs & receive camera kit.

03 // RECORD

Follow protocols and upload continuous takes.

04 // PASS QA

Pass privacy redaction & quality check.

05 // EARN

Receive weekly direct payouts per approved clip.

Minimum Age Required

18+ Years Old

Gear Package Status

Provided by Studio

Compensation Structure

Per Approved Episode

Approved Territories

International Openings

Ready to scope a custom dataset pilot?

Tell us the task family, environments, required annotations, and budget range — our studio team will follow up with a protocol proposal.