TrainThemAI/POV-Egocentric-Video-Robotics-FHD-Samples

---
license: mit
task_categories:
- video-classification
- robotics
tags:
- egocentric
- pov
- robotics
- human-demonstration
- vla
- manipulation
- activities-of-daily-living
pretty_name: TrainThem AI - Egocentric POV FHD Samples
size_categories:
- n<1K
---

# TrainThem AI — Egocentric POV FHD Samples

## Dataset Summary

This dataset contains high-fidelity, egocentric (first-person POV) video samples designed specifically for training generalist robotics policies (e.g., π0, π1, VLA models) and humanoid manipulation systems.

The data is collected by [TrainThem AI](https://trainthemai.com), a specialized agency providing diverse, rights-cleared human demonstration data at scale — coordinated via Slack, filmed with head-mounted cameras, and QA'd for robotics pre-training.

---

## Why Egocentric Video for Robot Training?

Robots trained to manipulate objects need to understand the world from the perspective of an embodied agent — not a third-person camera or a fixed overhead view. Head-mounted egocentric video provides:

- **The correct visual frame of reference** — the same POV a robot with a head-mounted camera experiences
- **Natural hand-eye coordination signal** — hands in frame throughout, matching the robot's own kinematic structure
- **Real-world distribution** — lighting variation, clutter, diverse objects and surfaces that sim-to-real transfer cannot replicate
- **Dense action supervision** — continuous manipulation sequences ideal for behavioral cloning and VLA fine-tuning

This is the data modality used by leading robotics labs for pre-training foundation models (e.g., RT-2, π0, RoboFlamingo, GR00T).

---

## Video Content

Three continuous activity episodes demonstrating Activities of Daily Living (ADL) and fine-motor manipulation:

| File | Activity | Focus |
|---|---|---|
| `tidying_the_bedroom.mp4` | Bedroom tidying | Object manipulation, spatial reasoning, organization |
| `organizing_bathroom.mp4` | Bathroom organization | Fine-motor sorting, varied materials, object placement |
| `washing_dishes.mp4` | Dish washing | Wet environment, bimanual coordination, tool use |

---

## Technical Specifications

| Property | Value |
|---|---|
| Resolution | 1920×1080 (FHD) |
| Frame rate | 30 fps |
| Perspective | Head-mounted egocentric (true POV) |
| Environment | Real-world domestic, non-simulated |
| Hands in frame | >90% of duration |
| Action density | No idle sequences >3 seconds |
| Rights | Fully cleared for AI training use |
| Total sample size | ~5.4 GB |

---

## QA Standards

All data from TrainThem AI meets the following criteria before delivery:

- True head-mounted POV (not handheld or chest-mounted)
- Hands consistently in frame during manipulation
- Action-dense — dead time removed
- Real domestic environments (not staged or simulated)
- Rights cleared — all collectors sign data licensing agreements

---

## Full Dataset & Bespoke Collection

These three clips are a limited evaluation sample. TrainThem AI currently operates at **150+ hours per day** of fully QA'd egocentric data capacity, covering:

- **100+ unique real-world environments**
- **80+ task subcategories** (kitchen, workshop, office, outdoor, etc.)
- Optional add-ons: 3D hand tracking, depth data (RGB-D), language annotations, multi-camera rigs

We offer both **off-the-shelf bulk data** and **bespoke collection** designed to your exact task taxonomy and annotation schema.

---

## Contact

To discuss bulk licensing, bespoke collection, or access to our full inventory:

- **Website:** [trainthemai.com](https://trainthemai.com)
- **Email:** [e-mail]

---

## Citation

If you use this data in your research or products, please cite: