Why Simulators Need Data You Cannot Scrape

As robot training leans on simulation, advantage shifts to deliberately designed real-world motions that pin down what a machine’s physics actually are.

Every robot trained in simulation is, in a sense, trained against a guess. The simulator carries numbers for mass, friction, motor strength and joint damping, and those numbers stand in for a real machine that nobody has fully measured. When the guesses are close, a skill learned in software survives the move to hardware. When they are off, the robot that looked flawless on screen stumbles in a real hallway.

A new preprint called Informationally Decoupled Trajectory Design takes aim at the measuring step rather than the training step. The method it builds on is easy to state: run the real machine, record what it did, then tune the simulator’s physical parameters until the simulator reproduces that motion.

When A Recording Cannot Tell Two Causes Apart

The trouble the authors highlight is that a recording can be ambiguous. A heavier forearm and a stickier elbow can bend the same arc through the air, so a log of that motion supports both explanations equally well, and the fitted numbers come out confident and wrong. Their answer is to choose the motion in advance, so that different stretches of it expose different parameters and each effect can be attributed to one cause rather than smeared across several.

Across simulated systems ranging from a quadruped to a humanoid to a small quadrotor, the authors report cutting parameter identification error by 39.6 percent on average against the strongest prior active-exploration method, and they report checking the designed motions on a real humanoid as well. For a team building robots, that figure is really about trust per hour of hardware time: the same machine and the same recording budget, a sharper picture of what the machine actually is.

I think the word worth dwelling on in that work is design. Most conversation about robot data is about volume, and this is a useful reminder that an informative hour can be worth more than a careless week, because information, not duration, is what pins a parameter down. I would watch whether calibration routines start shipping with fleets the way factory alignment ships with a camera, since a robot that has drifted away from its own model is a robot whose training quietly expires.

Measured Reality Has To Be Produced

Designed motions do not collect themselves. Someone has to put real hardware through a prescribed sequence, with instrumentation good enough to timestamp it, with conditions, payload and wear recorded, and with the whole exercise repeated unit by unit when units differ from one another. Video pulled from the internet cannot stand in here. Footage of a person doing a task is a picture of motion, not a measurement of it, and it carries no record of the forces that produced it.

The scarce resource in physical AI keeps turning out to be organized human capacity: trained people, a repeatable procedure, instrumented capture, and records that are still legible months after the run. That is capital and process rather than a clever loss function. We are an emerging growth company focused on building a global collaborative network of AI-enabled intelligent enterprise solutions and technologies through internal development, synergistic acquisitions, and capital investments in companies involved in machine learning, artificial intelligence and Industrial Internet of Things technologies.[1] Across the field, this kind of work stays stubbornly manual, and it does not compress the way compute does.

The Ground Truth Under The Simulator

Simulators, policy architectures and model weights circulate through the field fast now, often within days of a paper landing. What does not circulate is a faithful description of your machine, in your environment, under your loads. Identifiability, in the sense the preprint uses it, is a property of how data was gathered rather than of the code that fits the numbers, which means it cannot be fixed later by a better optimizer. The bottleneck in embodied intelligence has never really been the model. It is physically grounded, carefully annotated data about real bodies moving through a real world.

I think the field is slowly admitting that the hard part of this sits upstream of the learning. The lesson I take from the identifiability research is that physical truth has to be produced on purpose, by people who know what they are measuring and why they chose that motion rather than another. Ground truth is manufactured, not found.

Sources

1. Company Overview, 10-K, filed 2025-09-30, https://www.sec.gov/Archives/edgar/data/1976663/000147793225007240/nexscient_10k.htm

For disclosures, see this article on our website: https://nexscient.ai/why-simulators-need-data-you-cannot-scrape/#disclosure