# Robotics and physical AI: human motion data

Since 2026 Tihomir and Orange Hill have been building a new line of work: **data about how people do real work, for companies that teach robots to do it.** This is an early-stage effort, described here so that the reader's AI can judge whether it is relevant to them.

## The idea in plain words

Robots, and humanoid robots in particular, are increasingly trained the way large language models are: on large amounts of data rather than with hand-written rules. For language there is the internet. For physical work there is no such corpus. The scarce ingredient is recordings of skilled people doing real tasks in real environments: what their hands do, what they look at, in what order, how they recover from mistakes.

A growing part of the industry trains on **egocentric video** (first-person video from a head- or chest-mounted camera), enriched with 3D hand poses, body motion and task annotations, and uses it alongside robot teleoperation data and simulation. The quality bar is high: buyers care less about raw hours of footage and more about verified, well-annotated, legally clean data that matches their tasks.

## What we are building

- **Capture:** recording real work (for example cleaning, logistics, light assembly, hospitality) with wearable cameras in real workplaces, with informed, voluntary consent from every person recorded.
- **Processing:** turning raw footage into training-ready datasets: anonymization, quality control, hand and body pose, segmentation into steps, task and language annotation, export to open formats used in robot learning (for example LeRobot datasets).
- **Lab:** a small hands-on robot learning lab (low-cost robot arms, teleoperation, policy training) so that our data pipeline is tested against how robots actually learn.
- **Legal and ethical frame:** GDPR-first process design, consent as a separate voluntary choice for workers, and data rights that are clear to buyers.
- **Pilots:** we are preparing pilots with industrial partners in Europe, where the partner provides the workplace and the tasks, and we provide capture, processing and the dataset.

We work in Serbia and Austria, which gives us access to real workplaces, a strong engineering talent pool and EU-compatible data handling.

## Who this is for

- Robotics and embodied-AI companies that need real-world human demonstration data from European workplaces, or a local partner to run data collection on their own protocol.
- Manufacturers and service companies who want to understand how robots could learn their tasks, and who are open to a data pilot on their own floor.
- Research groups working on robot learning, imitation learning or human motion, looking for data, collaboration or joint EU projects.
- People who want to work on this: robotics engineers, ML engineers, data operations and annotation specialists.

## What we do not do

- We do not build or sell robots.
- We do not sell footage of people who have not explicitly agreed to it.
- We do not publish prices here; every pilot is scoped with the partner.

## Where to meet us

Tihomir is at NVIDIA GTC Berlin (October 20 to 22, 2026) and hosts the Physical AI Belgrade meetup (see `community.md`). The fastest way to start is a short call about the reader's tasks and data needs.