Asimov's human-data bet: crowdsourced videos train humanoids; YC W26 standout

A data-infrastructure company that pays a global network of contributors to record everyday human tasks and sells the resulting datasets to train humanoid robots.

Humanoids are data-bound, not hardware-bound: pay people worldwide to record everyday tasks, package it into training datasets, and own the supply chain first.

Anshul Verma (Scale AI, Amazon, Berkeley ML research) and Lyem Ningthou (US Air Force data pipelines, founding engineer at Blume, YC W24) - college roommates who previously co-founded a startup that reached six figures in revenue - founded Asimov in YC's Winter 2026 batch. Their bet: language models trained on trillions of words and image models on billions of photos, but robots are starting from scratch - the bottleneck is diverse real-world human movement data, and nobody owns the supply chain.

Asimov launched publicly in mid-March 2026, just before W26 Demo Day, and TechCrunch picked it as one of the 16 most interesting of the ~190-company cohort. By late March 2026, per AC Studio, more than 5,000 contributors worldwide were submitting videos of everyday tasks, with the company supplying thousands of hours of human-motion data daily to major robotics labs. Its August 2026 YC launch post said the largest robotics labs in the world were already customers, and that restaurant and warehouse partners can cover workers' salaries by having them wear collection kits while working.

Still running: contributor network live, datasets selling to major robotics labs; early-stage, three-person team.

The thesis is contrarian: while capital chases robot hardware and models, the bottleneck is upstream - nobody owns the diverse real-world movement data robots need to learn from.

The crowdsourcing model compounds: more contributors means more environments and tasks, making the dataset itself harder for a factory-capture competitor to replicate.

The founders are credible operators for the wedge: Scale AI taught Verma data infrastructure at scale, and Ningthou built USAF data pipelines and a previous startup to six figures in revenue.

Supply already beats demand signals: 5,000+ contributors within weeks of launch, thousands of hours of data flowing daily, and major labs as customers per the YC launch post.

When a category's hardware and models get the attention, the unglamorous input - training data - is the real bottleneck; crowdsourcing turns that scarcity into a defensible supply chain.

As of September 2026 Asimov runs a global contributor network and sells human-movement datasets to robotics companies, with a three-person team based in San Francisco. Its launch post says the largest robotics labs in the world are already customers and that it can deliver at scale immediately; it is recruiting contributors and business partners such as restaurants and warehouses to wear collection kits. The long bet is that robotics training data becomes a commodity input with a defensible, network-effect supply chain, the way Common Crawl became for text - still early and unproven.