Anthrobase / About

About

Building the data layer for Physical AI.

Anthrobase is a physical AI data infrastructure company in New Delhi. We capture how master practitioners see, decide, fail, and recover during real work, and structure it as training data for robotic foundation models.

Robot foundation models have most of the internet to learn from. What they lack is grounded human expertise.

The judgment, attention, and recovery of people who are masters of a physical task does not exist on the internet in any usable form. It lives in hands and habits, in how a line cook reads a pan or a mason reads a wall. Anthrobase exists to capture that, with consent, and turn it into a corpus models can actually learn from.

We are early. The first verticals are in active capture and the Expert Demonstration Standard releases as an open spec in 2026.

2026EDS open spec release
3Verticals in active capture
8Partner sites under MoU
New DelhiHeadquarters
What we do

Capture, structure, license.

Capture

We deploy synchronized sensor kits beside expert practitioners at partner sites and record six aligned streams of real work.

Structure

Every session is encoded under the Expert Demonstration Standard, reviewed by named technical reviewers, and tagged for tacit knowledge.

License

Datasets are licensed to robotics labs with explicit consent, frame-level provenance, and native exports to LeRobot, Open X-Embodiment, and RLDS.

What we believe

Principles we build on.

/ 01

Expertise over scale

A small corpus of true mastery is worth more than a large corpus of average work. We capture the best practitioners, not the most available.

/ 02

Consent is the product

People who give us their expertise are partners, not subjects. Consent, compensation, and the right to withdraw are built in, not bolted on.

/ 03

Built in India

India has extraordinary depth of physical skill across hospitality, craft, construction, and industry. We capture it where it lives, under Indian law.

/ 04

Open by default

The Expert Demonstration Standard releases as an open spec, with native exports to the formats the robotics community already uses.

Team

Started by two Carnegie Mellon graduates.

Anthrobase was founded by Umesh Sharma and Pranjal Chavarkar, who met as graduate students at Carnegie Mellon University. We are building a small, senior team in New Delhi across capture engineering, computer vision, and field operations.

Umesh Sharma

Founder & CEO

MS, Carnegie Mellon University, in Societal Computing and Machine Learning. Leads the capture systems and the Expert Demonstration Standard.

Pranjal Chavarkar

Co-founder

Carnegie Mellon University. Building Anthrobase's capture and data infrastructure.

Work with us, or join us.

Scope a dataset, become a capture site, or build the data layer for Physical AI with our team in New Delhi.