Build AI
"The data hyperscaler for Physical AI"
San Francisco data company collecting egocentric video of real factory workers; its open Egocentric-100K set holds 100,405 hours for robot learning.
Build AI is a San Francisco-based company, organised as a public benefit corporation, that produces large-scale egocentric (first-person) video datasets for training robot manipulation and other physical AI models. It describes itself as the data hyperscaler for Physical AI and says it is vertically integrated across capture hardware, manufacturing, logistics, data collection and model training. The company lists offices in San Francisco and Shenzhen and says it is launching collection operations in further countries. Its founders listed on the official site are Patrick Rim, Eddy Xu, Jonathan Jia and Zikang Jiang; Eddy Xu leads the company and Patrick Rim is chief scientist.
The core product is footage recorded with Build AI's own monocular head-mounted capture devices (the Build AI Gen 1 device is named in the Egocentric-100K release) worn by workers in real factories and businesses while they perform skilled manual tasks. Because the footage shows hands interacting with tools, parts and fixtures from the worker's viewpoint, it is aimed at research groups and companies that train vision-language-action models, humanoid manipulation policies and world models from human demonstration video rather than only from robot teleoperation.
Build AI has released its datasets in a scaling series. Egocentric-10K, published on Hugging Face in August 2025 under the Apache 2.0 licence, contains 10,000 hours and 1.08 billion frames of 1080p, 30 fps video, organised by factory (folders run to factory_085) and worker, with an average of 4.68 hours per worker. Egocentric-100K followed in December 2025 with 100,405 hours of footage from 14,228 factory workers, about 10.8 billion frames and 24.79 TB, released as 256p, 30 fps H.265 video. The official site lists Egocentric-1M for March 2026 and roadmap targets of Egocentric-10M (November 2026), Egocentric-100M (October 2027) and Egocentric-1B (December 2028). Its commercial off-the-shelf dataset is captured at 1920x1080 and 30 fps, and the company offers service-level guarantees on 3D hand and body pose accuracy (MPJPE), diversity and quality. Access to the Hugging Face releases is gated behind a contact-information agreement.
On funding, industry press reported a 5 million US dollar seed round led by Abstract, Pear and HF0, later updated to 15 million dollars with angels including Balaji Srinivasan, Guillermo Rauch and Thomas Wolf; the official site now states that the company has raised more than 30 million dollars, naming HF0, Abstract, Pear, Renegade, Costanoa, Nomad, Deepwater, Cyber and SF100 among its backers. Industry overviews of egocentric data providers for robotics, such as one published by Labellerr, list Build AI among the main suppliers of human-demonstration data for embodied AI. For RobotToday the company sits in the data, teleoperation and simulation segment of the humanoid and embodied-AI supply chain as a provider of training data rather than robots.
Primary type & automation activities this supplier delivers:
Egocentric-10K
Open Apache-2.0 dataset of 10,000 hours of head-mounted first-person video from factory workers for robot manipulation and physical AI training.
Egocentric-100K
Open egocentric dataset of 100,405 hours from 14,228 factory workers captured on Build AI Gen 1 head-mounted cameras.
Egocentric-1M
Third dataset in Build AI's scaling series, listed on the official timeline for March 2026; commercial data offered with SLAs on 3D pose MPJPE, diversity and quality.
More related news
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
Contact Build AI
WEBSITE
https://build.aiHEADQUARTERS
United States
Company Facts
Founded
2025
Primary Role
Software/Algorithm
Company Size
-
Primary Region
North America
Annual Sales
-
Funding Stage
Early-Stage VC
Funding Total
-
