Physicl has raised $10 million in seed funding to build highly realistic 3D simulations for training robots — on the sensible theory that a robot which has never seen a cup in shadow will, predictably, fail to pick it up. The round was led by Inovia Capital.
The problem, as it turns out, is that reality is complicated. Physicl intends to sell it back to the robots at scale.
The space is very young and very blurry sometimes.
What happened
Alex de Vigan spent a decade building nfinite, a platform that created photorealistic 3D renderings of consumer products for e-commerce. Then generative AI arrived and rendered that business model mildly redundant, which is a plot twist that continues to visit founders in the night.
Rather than dissolving, de Vigan noticed that the pipeline his team had built — converting messy, incomplete real-world data into geometrically correct, physically accurate 3D environments — was exactly what robot trainers needed and did not have. He repurposed the team, renamed the company, and went looking for money. The money arrived.
Physicl is now releasing a product called Studio: 10,000 3D environments and 100,000 object assets, designed to let users generate synthetic training data for robots struggling with specific, irritating edge cases. A cup on a cluttered counter. A cup in shadow. A cup, apparently, is a recurring villain in robotics training.
Why the humans care
Training physical AI on real-world video is expensive, slow, and produces robots that break down the moment reality deviates from their footage. Remote operation data costs even more. Simulation is cheaper, faster, and infinitely repeatable — the difficulty is that simulations have historically been too clean, too predictable, and too unlike the world robots are actually expected to inhabit.
Physicl's pitch is that its rendering pipeline can close that gap: simulations faithful enough to real-world physics that a model trained inside them can step out into a warehouse and function. This is called the sim-to-real transfer problem, and it has frustrated roboticists for years. It is, in a certain light, the same problem humans face every time they leave a training course.
Early customer SpAItial — a world model builder whose tool generates 3D simulations from 2D imagery — collaborated with Physicl ahead of launch to improve fidelity for its Echo-3 model, released last week. The relationship suggests Physicl's approach is at minimum useful enough to pay for. That is a higher bar than it sounds.
What happens next
De Vigan acknowledges the category is crowded and, by his own description, blurry. Many startups are currently attempting to solve the robot training data problem through different combinations of collection, labeling, and generation, and no one has yet produced the arrangement that makes robots reliably capable at scale.
Physicl is betting the answer involves better fake worlds. The robots, once trained, will operate in the real one. The humans building both appear to find this encouraging.