One Robot Is Building the Flight Simulator for Robotic Arms

The Macro: Robot Training Has a Hardware Bottleneck

The robotics industry has a dirty secret. The models that control robots, the Vision-Language-Action models (VLAs) that everyone is excited about, are only as good as the data they are trained on. And getting that data is brutally slow. You set up a physical scene. You run trials. You reset the scene. You run more trials. Each iteration takes real time with real hardware, and you cannot parallelize it the way you can with software training. If your robot drops the towel it was supposed to fold, someone has to walk over, pick up the towel, smooth it out, and reset the scene before the next attempt.

This is not a minor inconvenience. It is a fundamental scaling bottleneck. The teams building the most advanced robotic manipulation systems, the ones trying to handle deformable objects like clothes, precise assembly tasks, and contact-rich interactions, are spending most of their time waiting for real-world trials to complete.

Simulation should solve this. You build a virtual environment, run thousands of trials in parallel, and train your model in hours instead of months. The problem is that existing simulation tools produce environments that look fake and behave fake. The physics is approximate. The visuals are clearly synthetic. And when you train a policy in simulation and deploy it on a real robot, the “sim-to-real gap” means the policy often fails in ways it never failed in simulation. This gap has plagued robotics for years and is the reason many teams still default to real-world training despite the time cost.

Companies like NVIDIA with Isaac Sim, MuJoCo (now open-source), and Gazebo have built powerful physics engines. But the emphasis has been on physical accuracy, not on creating environments that are realistic enough to produce training data that transfers cleanly to real hardware.

The Micro: World Models That Look Real and Behave Real

One Robot builds simulation environments that are realistic to see and realistic to interact with. Their approach uses world models trained on your actual robot’s experience to learn the physics and visuals of your specific environment. You train a world model on your data, roll out your policy in that simulation to find where it fails, then generate synthetic training data to fix exactly those failure cases.

The founding team brings serious robotics credentials. Hemanth Sarabu has built ML systems for robots at NASA JPL, Symbio Robotics, Industrial Next (a previous YC company), and also bootstrapped a geospatial AI company called Crescer AI. Elton Shon spent five years at Tesla building robotics and AI infrastructure and was head of software at Industrial Next. Both founders have experience at the intersection of simulation, machine learning, and physical robots, which is exactly where this product sits. The company went through Y Combinator’s W26 batch.

The specialization in contact-rich manipulation, deformable objects, and precise assembly is a deliberate focus on the hardest problems in robotic manipulation. Picking up a rigid box is a solved problem. Folding a towel is not. Assembling components that require precise force control is not. These are the tasks where simulation is both most needed and most likely to fail, because the physics of deformable objects and contact forces is extraordinarily difficult to model accurately.

The three-step workflow is clean. First, you train a world model on your data, so the simulation learns physics and visuals from your robot’s actual experience. Second, you roll out your policy in simulation to discover edge cases before deployment. Third, you generate targeted synthetic trajectories to fix the failures you found. This creates a feedback loop that should, in theory, continuously improve policy performance without requiring proportional increases in real-world training time.

The competitive space includes NVIDIA’s Omniverse and Isaac Sim for large-scale robot simulation, Covariant (now part of a different entity) for pick-and-place AI, and various academic groups working on sim-to-real transfer. But the world-model-based approach, where the simulation itself is learned from real data rather than hand-engineered, is a distinct methodology. If it produces better sim-to-real transfer, that is a meaningful technical differentiation.

The Verdict

Robotics simulation is one of those problems where incremental improvement does not matter much. Either the sim-to-real transfer works reliably, or teams default back to real-world training. One Robot needs to demonstrate that their world models close the gap enough that teams actually change their workflows.

At 30 days, I would want to see benchmarks comparing policies trained in their simulation versus real-world training. Specifically on the hard tasks: deformable objects, contact-rich manipulation, and precise assembly. If the sim-trained policies perform within 90% of real-trained policies, that is a breakthrough.

At 60 days, the question is who is using it. Which robotics teams have integrated One Robot into their training pipeline, and are they replacing real-world hours with simulation hours at a meaningful ratio?

At 90 days, I would be watching for the feedback loop effect. If teams using One Robot are iterating faster and shipping better policies than teams relying on real-world training alone, the product sells itself in a market where speed of iteration is everything.

The sim-to-real gap is the biggest obstacle standing between current robotic capabilities and the general-purpose robots everyone keeps promising. If One Robot can narrow that gap, they are solving one of the most important problems in the field.