---
license: cc-by-4.0
pretty_name: >-
  fidelity-data-factory/ README.md Metadata UI license   task_categories 
  language  tags  pretty_name  size_categories
size_categories:
- 100K<n<1M
task_categories:
  - reinforcement-learning
  - representation-learning
  - video-understanding

tags:
  - egocentric
  - robotics
  - state-action-state
  - world-models
  - vla
  - human-interaction
  - imitation-learning
  - offline-rl

---


# Fidelity Data Factory – Egocentric State–Action Transitions (v0)

This repository contains an initial release of structured state–action–state′
transitions extracted from real-world egocentric video.

The goal of this dataset is to provide early infrastructure for learning
dynamics and representations from large-scale human activity data.

## Overview

Each data point is a short temporal transition of the form:

(s_t, a_t, s_{t+1})

Transitions are derived from monocular egocentric footage recorded in real
factory environments.

This release does not include robot-specific signals such as torques or joint
states, and is intended for research and exploration rather than deployment.

## Data Contents

- ~200k+ transitions
- Egocentric (head / chest-mounted) viewpoint
- Real industrial environments

Transitions are stored in JSONL format.

## Schema (Simplified)

Each record contains:

- `s`:
  - `ego_pose`
  - `ego_velocity`
  - `hand_state`
  - `entities` (objects with image-space location)
  - `meta` (video id, timestamp)

- `a`:
  - `ego_delta`
  - `hand_delta`
  - `interaction_delta`

- `s_prime`:
  - Same structure as `s`, representing the next timestep

See `schema.json` for full details.

## Intended Use

This dataset may be useful for:
- World model research
- Offline RL
- Vision–language–action pretraining
- Learning dynamics from human activity
- Representation learning from egocentric video

## Limitations

- Monocular video only
- No force / torque signals
- No task labels
- Contains estimation noise

## Credits

Original video data provided by BuildAI.  
Enrichment and processing by Fidelity Dynamics.