Exocentric Data Collection for Physical AI

See the full picture. Train for the real world.

Sourcebae captures third-person video of people performing real tasks across homes, factories, warehouses, workshops, and commercial environments.

Using external cameras, we record human movement, object interactions, and complete workflows to build training data for robotics, embodied AI, and world models.

Third-person capture of people, objects and environments feeding a Physical AI model
External cameras tracking people as they move through a kitchen, warehouse and workshop

Real-world actions. A wider perspective.

Robots need to understand more than what a person sees.

They need to understand how people move through spaces, interact with objects, use tools, and complete tasks within their surroundings.

Exocentric data captures these activities from an external point of view, providing a wider view of the person, workspace, and environment.

  • People
  • Objects
  • Movement
  • Space
  • Actions

What is Exocentric Data Collection?

Exocentric data collection records human activities using cameras positioned outside the person performing the task.

Unlike egocentric video, which captures the first-person perspective, exocentric video shows the person and their surrounding environment from a third-person view.

This makes it useful for understanding body movement, spatial relationships, object interactions, and how tasks unfold over time.

Sourcebae designs custom exocentric collection programs based on your model's tasks, environments, camera requirements, and data specifications.

Multiple external cameras covering a workshop as a person works at a bench

Capture real tasks from the outside

Human Movement

Record walking, reaching, bending, lifting, carrying, and other natural body movements.

Object Interactions

Capture how people pick, place, move, use, and interact with objects in real environments.

Complete Workflows

Record multi-step tasks from beginning to end, including task order, transitions, and outcomes.

Spatial Context

Understand how people move around furniture, equipment, tools, and other objects within a workspace.

Success & Recovery

When required, capture failures, fixes, recoveries, and successful demonstrations.

One network. Multiple real-world environments.

We collect exocentric data across different industries and everyday environments.

Household task recorded with external cameras

Household

Cooking, cleaning, organizing, folding clothes, appliance use, and everyday movement.

Warehouses & Logistics task recorded with external cameras

Warehouses & Logistics

Picking, placing, sorting, packing, loading, and movement through workspaces.

Retail & Commercial task recorded with external cameras

Retail & Commercial

Shelf stocking, inventory handling, commercial kitchens, housekeeping, and store operations.

Manufacturing task recorded with external cameras

Manufacturing

Assembly, inspection, machine operation, packaging, and material handling.

Workshops & Automotive task recorded with external cameras

Workshops & Automotive

Mechanical repair, equipment servicing, tool use, and vehicle maintenance.

Agriculture & Outdoor task recorded with external cameras

Agriculture & Outdoor

Harvesting, sorting, grading, tool use, and outdoor work activities.

Fixed, tripod-mounted and overhead cameras arranged around a workbench

Capture setup built around your model

Every project has different requirements for camera placement, coverage, and data quality. Sourcebae can support custom setups using:

  • Fixed external cameras
  • Tripod-mounted cameras
  • Multiple synchronized camera views
  • Wide-angle video capture
  • RGB / RGB-D cameras
  • Optional audio and sensor signals

Camera positioning, resolution, frame rate, and synchronization requirements are defined during project scoping.

When to Use Exocentric Data Collection

Use exocentric data when your model needs to understand more than what the person performing the task can see.

  • Whole-Body Motion captured from an external camera

    Whole-Body Motion

    Capture walking, reaching, bending, lifting, posture and coordinated movement from a stable external viewpoint.

  • Spatial & Workspace Context captured from an external camera

    Spatial & Workspace Context

    Understand where people, tools, equipment and objects are located and how they move through a shared environment.

  • Multi-Person & Human-Robot Interaction captured from an external camera

    Multi-Person & Human-Robot Interaction

    Record interactions between people, robots, tools, machinery and moving objects within the same scene.

  • Task Outcomes & Recovery captured from an external camera

    Task Outcomes & Recovery

    Observe successful actions, failed attempts, corrections and final outcomes from an independent viewpoint.

More than third-
person video

Your dataset can be configured to include the information your training pipeline needs.

  • Video Recordings

    External-view footage of people performing real tasks.

  • Multi-View Capture

    Synchronized recordings from multiple angles where required.

  • Action & Task Labels

    Task steps, action segments, object interactions, and task outcomes.

  • Human Motion Data

    Movement, pose, or trajectory annotations per project needs.

  • Metadata

    Task, environment, camera, contributor, and recording information.

  • Custom Delivery

    Files, annotations, and datasets aligned with your specs.

Built for Physical AI and robotics

  • Human Motion Understanding

    Learn how people move and perform physical activities.

  • Robot Imitation Learning

    Use human demonstrations to support the learning of real-world tasks.

  • Manipulation & Object Interaction

    Understand how people handle tools, materials, and objects.

  • World Models

    Capture how actions change the state of objects and environments.

  • Spatial Reasoning

    Learn relationships between people, objects, and surrounding spaces.

  • Robotics Evaluation

    Evaluate models across different tasks, environments, and real-world conditions.

Sourcebae contributors including factory workers, technicians, healthcare staff, lab scientists and household contributors

The right people. The right environment.

Not every task requires the same contributor.

Sourcebae can source people based on the skills and experience needed for your collection.

From household contributors to factory workers, technicians, and specialized professionals, we help capture tasks performed by people who understand the work.

Real skills create more meaningful demonstrations.

From your requirements to
training-ready data.

  • 01 - DEFINE

    We understand your target tasks, environments, camera setup, and quality requirements.

  • 02 - SOURCE & SET UP

    We arrange contributors, locations, equipment, and recording protocols.

  • 03 - PILOT

    A small collection validates camera placement, task instructions, and output quality.

  • 04 - CAPTURE

    Contributors perform real tasks while external cameras record each setup.

  • 05 - QA & ANNOTATE

    Recordings are checked for visibility, task completion, and quality standards.

  • 06 - DELIVER

    Approved data is organized and delivered in the agreed format.

Exocentric vs Egocentric
Data Collection

FeatureExocentricEgocentric
PerspectiveThird-person / externalFirst-person
CameraFixed or externalWorn by participant
Best visibilityBody, environment, spatial contextHands, tools and close interactions
Multi-person activityStrongLimited
Workspace geometryStrongPartial
Human intent / POVIndirectStrong
Typical robotics useMotion, spatial reasoning, task evaluationManipulation and imitation learning

For some Physical AI programs, synchronized egocentric and exocentric capture provides both the actor's first-person view and the wider scene context around the task.

Quality starts before
recording

Good exocentric data requires more than a camera pointed at a task. Sourcebae builds quality checks into the collection process.

  • Capture Quality

    • Camera position
    • Lighting
    • Task completion
    • Occlusion
    • Recording continuity
    • Multi-view coverage
  • Dataset Quality

    • Metadata validation
    • Annotation checks
    • File integrity
    • Synchronization checks
    • Recapture when required
  • Privacy & Governance

    • Participant consent
    • Site authorization
    • Privacy boundaries
    • Sensitive-information handling
    • Face/identifier redaction when required
    • Secure dataset delivery

Why Sourcebae?

  • Real-World Access

    Collect across household, commercial, industrial, and specialized environments.

  • Skilled Contributors

    Match participants to the tasks your model needs to learn.

  • Flexible Camera Setups

    Configure external cameras, multiple views, and additional sensors around your requirements.

  • End-to-End Execution

    Contributor sourcing, setup, capture, QA, annotation, and delivery managed through one workflow.

  • Pilot Before Scale

    Validate the data first, refine the protocol, and then expand production.

Frequently asked questions

  • Exocentric data collection captures activities from an external or third-person camera perspective. It records people, objects, movement, and the surrounding environment during real-world tasks.

Build a wider view of the physical world

Tell us what your model needs to learn. We'll help define the people, environments, camera setup, and quality workflow needed to capture it.