World models for VLA evals and training.
TL;DR
One Robot builds world model-based simulations that are realistic to see and realistic to interact with, so robotics teams can train and evaluate robot action models (VLA) without being bottlenecked by robot time.
The Problem
Improving a VLA today often means more real-world hours: setting up the scene, running trials, resetting, and repeating. This loop is slow, expensive, and fundamentally unscalable. To generalize, models need far more training and evaluation data than teams can realistically collect in the real world.
Hemanth and Elton used to train robots to automate the assembly of parts requiring sub-millimeter precision. As models got better, they hit a surprising bottleneck: it started taking longer to evaluate a new policy than to train it. To answer “Is model B better than model A?”, they had to:
Schedule a physical robot → run tests → reset scene → record failures → tweak → retrain models → repeat.
Simulations should solve this. But existing simulators fall short because they’re either not photorealistic or not physics-realistic. Neither is sufficient for learning and evaluating modern robot policies.
The Solution
We build task-specific world models that learn contact dynamics and visual appearance from robot data, and use them to generate realistic simulation environments for training and evaluation. Instead of relying on generic physics engines, our models predict how objects move and deform under robot actions, achieving photo- and physics-realistic rollouts for hard manipulation tasks like textiles and box folding. These environments help teams run more training and evals, find failure modes faster, and accelerate iteration on action policies with less dependence on real-world data collection and robot availability.
The Team
Hemanth built ML systems for robots at Google, NASA JPL, Symbio Robotics, and Industrial Next (YC W22). He previously bootstrapped a profitable geospatial AI company and built vehicle simulations at McLaren.
Elton spent five years at Tesla, building robotics and large-scale AI infrastructure for factory automation during the Model 3 production ramp and the Dojo supercomputer team. He later joined Industrial Next (YC W22) as head of software, building robot learning and control systems.
Ask
If you or someone you know is training robots for real-world tasks with manipulators, let’s connect! (founders@onerobot.io)
