<
Updated 2026
First person video data

Egocentric video datasets for Physical AI, robotics, and multimodal models

Robots and embodied AI systems experience the world from inside the action. Egocentric video captures that first person perspective: hands interacting with objects, tasks unfolding over time, changing environments, and the spoken context that can explain what the camera sees.

Pangeanic provides commercially licensable egocentric video datasets spanning thousands of hours across Latin America and Asia. Its current inventory includes 1,500+ hours captured across Latin America and 5,000+ hours captured in Asian manufacturing environments, while custom programs can target specific tasks, geographies, participants, devices, and model requirements.

This capability connects our broader video dataset services , off the shelf training data , geographically targeted collection , and governed AI Data Operations .

What buyers can source

From existing egocentric footage to purpose built collection

Existing inventory

Pangeanic provides more than 6,500 hours of commercially licensable egocentric video across Latin America and Asia. The inventory includes 1,500+ hours captured across Latin America and 5,000+ hours captured in Asian manufacturing environments, with narration coverage, regional metadata, technical specifications, and licensing terms confirmed per procurement.

Narrated video

Collections can include synchronized spoken narration that describes actions, decisions, objects, and context as tasks are performed.

Custom programs

New data collection designed around the target model, intended deployment conditions, tasks, locations, devices, languages, and quality thresholds.

Procurement note: availability, permitted uses, jurisdictions, task coverage, and technical characteristics are validated against the current inventory before commercial commitment.

Why egocentric video

The open web shows the action from outside. Physical AI must understand it from within.

Egocentric video reduces the perspective gap faced by models that will operate from a head, wrist, gripper, wearable device, or mobile camera. Third person video is abundant, but abundance can be a mirage when it presents the action from the wrong viewpoint. First person data preserves the sequence between perception, intention, movement, and result.

01 · Viewpoint

Data closer to model deployment

A moving first person camera records the scene from a perspective closer to the one encountered by robots, wearables, and embodied systems.

02 · Interaction

Hands, objects, tools, and motion

First person footage can capture how people reach, grasp, manipulate, assemble, inspect, and move through real environments.

03 · Temporal context

Actions that unfold as sequences

Longer recordings preserve preparation, intermediate steps, corrections, interruptions, and outcomes that isolated frames simply discard.

04 · Language

Narration adds a semantic layer

Synchronized narration can connect visible actions with human intent, terminology, decision points, and culturally specific descriptions.

ECO Intelligence Platform: The operational platform and execution layer for sovereign multilingual AI
Egocentric video procurement

Start with available footage or collect for the exact failure mode

Pangeanic provides two procurement routes for egocentric video data: existing licensed inventory and custom collection. Some programs need data immediately, while others require a particular combination of task, environment, viewpoint, participant, language, and device. One generic corpus rarely fits every robot under the sun.

Existing inventory

Commercially licensable egocentric video

Pangeanic provides existing egocentric video inventory comprising 1,500+ hours captured across Latin America and 5,000+ hours captured in Asian manufacturing environments. Narration coverage, regional metadata, task distribution, technical specifications, and licensing terms are confirmed per procurement and dataset subset.

Procurement package can include:

Representative samples · Inventory summary · Technical specifications · Available metadata · Narration details · Delivery schedule · Commercial AI and ML licensing terms

Request current inventory
Custom collection

Purpose built data for the target system

Custom programs are scoped around what the model must perceive, predict, or perform. Collection protocols can specify the task ontology, recording setup, participant profile, geography, environment, narration, metadata, quality gates, and consent requirements.

Program design can cover:

Pilot collection · Scenario design · Contributor recruitment · Capture instructions · Automated and human quality control · Recollection rules · Annotation · Secure delivery

Explore geographically targeted collection
Dataset design

The recording is only the visible layer

Pangeanic structures egocentric video datasets around explicit capture protocols and acceptance rules. Each protocol defines what the participant must do, what must remain visible, how narration is captured, which metadata travels with each file, and what causes a recording to be rejected or repeated.

Task ontology

Activities with defined boundaries

Tasks, subtasks, start and end conditions, objects, tools, expected outcomes, and permitted variations are defined before collection begins.

Capture conditions

Viewpoint, device, and environment

Programs can control camera position, orientation, resolution, frame rate, lighting, movement, visibility, environment, and recording duration.

Audio and narration

Speech synchronized with action

Narration may describe the task as it unfolds or explain it afterward, depending on whether the model needs concurrent language, cleaner exposition, or both.

Metadata

Context that makes footage usable

Structured fields can cover task, location, language, regional variety, device, participant attributes, environment, duration, objects, and validation status.

Annotation

Training signals aligned over time

Optional enrichment can include clip segmentation, temporal action labels, object references, transcriptions, captions, event boundaries, and reviewer decisions.

Quality control

Acceptance rules before scale

A pilot calibrates technical checks and human review before wider collection, with explicit rejection, recollection, escalation, and delivery rules.

Model applications

Egocentric data for systems learning to perceive and act

Egocentric video supports Physical AI, robotics, world models, video language models, wearable systems, and activity recognition. Video alone may be sufficient for one model, while another needs narration, temporal labels, synchronized sensors, or a collection protocol matched to a specific embodiment.

Robotics

Robotics and manipulation

Human demonstrations involving grasping, sorting, assembly, tool use, inspection, household tasks, and industrial operations.

Embodied intelligence

World models and embodied AI

Longer sequences that connect visual observations, human actions, environmental change, and task outcomes.

Multimodal training

Multimodal and video language models

Video paired with narration, transcription, captions, instructions, or structured descriptions of actions and intent.

Wearable systems

AR, VR, and wearable systems

First person scene understanding, contextual assistance, user interaction, and activity recognition from body mounted devices.

Operational environments

Industrial and field operations

Procedural tasks, maintenance, logistics, inspection, safety workflows, and domain specific use of tools and equipment.

Contextual assistance

Activity recognition and assistance

Models that identify activities, understand task progression, detect deviations, or provide contextual guidance.

Rights, privacy, and delivery

A video file becomes an AI asset when its use can be explained and defended

Pangeanic documents provenance, permitted use, licensing, and privacy controls for commercially supplied egocentric video data. First person footage can reveal participants, bystanders, homes, workplaces, screens, documents, voices, and location cues, so commercial usability depends on more than visual quality.

Pangeanic structures these requirements within broader AI Data Operations so procurement, technical, legal, and model teams can assess the same asset without playing a small game of evidential archaeology.

Provenance and participant consent

Collection origin, participant permissions, and relevant usage rights are documented according to the agreed procurement and licensing framework.

Privacy aware preparation

Review, filtering, exclusion, and masking workflows can be applied where faces, screens, documents, identifiers, or sensitive surroundings require additional control.

Commercial AI and ML licensing

Permitted uses, restrictions, term, exclusivity, derivatives, model training rights, and redistribution conditions are defined for the selected dataset or collection program.

Structured and controlled delivery

Files, metadata, validation results, documentation, and delivery manifests are prepared according to the agreed schema and secure transfer process.

Related pathways

Connect egocentric video with the wider data operation

Pangeanic connects egocentric video sourcing with regional coverage, annotation, speech, privacy review, evaluation, and governed delivery. This creates a coherent multimodal data operation rather than a set of disconnected files and vendors.

Parent capability

Video datasets

General video sourcing, preparation, temporal annotation, metadata, and multimodal workflows.
Commercial path

Off the shelf training data

Existing licensed assets for programs that do not need to begin with a new collection.
Regional collection

Geographically targeted data collection

Data programs designed around authentic geography, language, culture, and deployment conditions.
Annotation

PECAT multimodal workflows

Controlled annotation, validation, multilingual review, quality assurance, and traceable delivery.
Audio

Speech datasets

Native speech, transcription, regional language coverage, audio validation, and conversational data.
Privacy

Data masking

Privacy aware preparation for visual, textual, and multimodal data containing sensitive information.

Frequently Asked Questions

Egocentric video datasets FAQ

What is an egocentric video dataset?

An egocentric video dataset contains footage recorded from the first person perspective of someone performing an activity. The camera may be worn on the head, chest, wrist, or another body position, or mounted on a device or robot. The footage captures the scene from inside the action rather than from a fixed external viewpoint.

How is egocentric video different from ordinary video training data?

Ordinary video is often recorded by a fixed or external camera. Egocentric footage moves with the person or system performing the task and captures first person attention, hand and object interaction, occlusion, motion, and task progression. This can better represent the visual perspective encountered by robotics, wearable, and embodied AI systems.

Does Pangeanic offer existing egocentric video inventory?

Yes. Pangeanic provides 1,500+ hours of egocentric video captured across Latin America and 5,000+ hours captured in Asian manufacturing environments. Recording format, narration coverage, metadata, licensing, and delivery timing are confirmed per procurement.

Can egocentric video include synchronized narration?

Yes. Narrated first person video can pair the visual sequence with a spoken description of actions, objects, decisions, and context. Narration may be recorded during the task or added afterward. Its availability and synchronization should be verified for each existing dataset or specified in a custom program.

Can Pangeanic collect custom egocentric video?

Yes. Custom collection can be designed around the target activities, locations, participant profiles, devices, languages, narration requirements, environments, technical specifications, metadata, consent, and quality criteria. A pilot is normally used to validate the protocol before scaling.

What annotations can be added to egocentric video?

Depending on the model and project, enrichment may include temporal segmentation, action labels, task and subtask boundaries, object references, transcriptions, captions, event descriptions, visible hand or tool status, quality flags, and structured metadata. Not every annotation type is included in existing inventory by default.

How are licensing, consent, and privacy handled?

Pangeanic defines commercial usage rights, restrictions, consent requirements, provenance documentation, privacy controls, and delivery conditions for the selected inventory or custom program. Where required, footage can also pass through review, exclusion, or masking workflows before delivery.

Which AI systems use egocentric video data?

Egocentric video can support robotics, embodied AI, Physical AI, world models, video language models, activity recognition, industrial assistance, AR and VR, wearable systems, and models that must understand human interaction with objects and environments over time.

Talk to Pangeanic

Need egocentric video for a model that must understand the physical world?

Tell us the target use case, required volume, tasks, locations, devices, narration, metadata, licensing needs, and delivery schedule. We will identify whether existing inventory fits or whether a custom collection will produce the stronger training asset.

Request samples and specifications Explore existing egocentric datasets Explore AI Data Operations