Kailash Labs is a real-world robot evaluation lab. We deploy general-purpose robot models on real robots doing economically useful work, measure where they succeed and where they break, and turn those failures into the experience these models are missing. Towards a world where anyone with any robot and any task can pick a capable model, collect a handful of demos, and get a policy they can trust.
What we’re betting on
01.Generalist models are the new default.
VLAs and frontier multimodal models are scaling the way language models did. The era of one specialist model per task is ending: a capable base model, a few demonstrations, and a robust policy for any robot and any task.
02.Nobody knows what these models can actually do.
Demos are cherry-picked, simulation benchmarks saturate, and lab results don’t survive contact with a real workplace. Before anyone can deploy, someone has to measure honestly: many tasks, many environments, many trials, with a careful record of how things fail.
03.Physical reality is the only benchmark that counts.
We run policies in a real robot workplace on tasks people would pay for. Every failure and every human recovery becomes structured data: exactly the real-world experience generalist models lack.
The trusted measure of robot capability will become the reference everyone deploying robots relies on. We intend to build it.
If you are training robot models or deciding which one to deploy, we want to hear from you.