How a Failed E-Commerce Tool Found a Second Life Training Robots

Alex de Vigan spent a decade building software that generated tempting product renderings for online retailers until generative AI essentially swamped his company's core market. Instead of closing shop, he realized the underlying engine designed to map unstructured imagery into accurate three-dimensional shapes solved a completely different bottleneck elsewhere.

That realization birthed Physicl, which has now secured funding from Inovia Capital to help robot builders escape a stubborn training trap. Robot developers aiming to make machines handle delicate, everyday tasks find themselves frustrated by the limits of standard video training and the high cost of manual remote operations.

Simulating digital worlds offers a promising escape hatch, but teaching machines physics that actually match reality remains notoriously difficult. Physicl steps into this crowded category with a product called Studio, packed with thousands of virtual assets built to generate synthetic training data for specific edge cases, like spotting a cup sitting in a shadow on a cluttered counter.

Our read is that the race to build physical AI is quietly shifting away from raw data collection toward automated curation, turning old software stacks into critical industrial infrastructure.

Early customer Matthias Niessner of SpAItial notes that instead of manually curating training sets over long cycles, tools can now auto-generate permutations on the fly depending on a model's current needs.

Whether this automated pipeline will finally make robots reliable outside controlled laboratory settings remains the ultimate test for the startup.