Industries / Industrial equipment and services / IoT and industrial automation / Digital twins
Simulation
Simulation is the part of a twin that calculates what the real thing will do next, from physics, from past data or from both.
4
Milestones achieved, from 1991 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1991
Model error sinks a platform
On 23 August 1991 the concrete base of the Sleipner A platform sank during a test in which it was lowered into the water. Statoil's investigators found that the finite element analysis, a calculation that splits a structure into many small pieces (23,000 in this case), put the shear forces in the walls at "only 53% of the correct values", and that the joint where three cells met "was not designed separately".
LessonA large, detailed model can still be wrong by half, so its results need an independent check before anything is built on them.
Credit
Statoil's investigation, reported by Oil & Gas Journal on 28 October 1991
1997
A shared language for models
Modelica is a language for describing physical systems, mechanical, electrical and thermal, as equations, so that models from different tools can be combined. The Modelica Association says the design effort "was initiated by Hilding Elmqvist, Dynasim AB, and started in September 1996", and that "a first version of a Modelica language description was put on the web in September 1997".
Credit
Hilding Elmqvist; the Modelica Association
2010
Models packed to travel
The Functional Mock-up Interface (FMI) is "a free standard that defines a container and an interface to exchange dynamic simulation models", so that a model built in one tool can run inside another. Version 1.0, for model exchange, is dated 26 January 2010. It came from MODELISAR, a European ITEA2 project that ran from July 2008 to December 2011 and was co-organised by Dassault Systèmes and Daimler. The FMI website now says it is supported by "290+ tools".
Credit
the MODELISAR partners; the Modelica Association, which now maintains FMI
2024
A national review of gaps
The US National Academies of Sciences, Engineering, and Medicine, for sponsors including the Department of Defense, the Department of Energy, the National Institutes of Health and the National Science Foundation, published Foundational Research Gaps and Future Directions for Digital Twins. It says "The bidirectional interaction between the virtual and the physical is central to the digital twin", and that a twin "goes beyond simulation". It concludes that "there is a lack of standards in reporting VVUQ as well as a lack of consideration of confidence in modeling outputs". VVUQ means verification, validation and uncertainty quantification: checking that a model is built right, that it matches reality and how sure it is.
nationalacademies.orgnationalacademies.orgnationalacademies.org
Credit
the National Academies' committee on digital twins
Watch for
World models as simulators
Google DeepMind calls Genie 4 a "new world model", a program that draws an interactive scene and changes it in answer to what the user does, and offers it for "safe, infinitely variable simulations", with examples of training robots such as a UR5 arm pipetting. It also says Genie 4 "remains an early, experimental model" whose "visual details and object permanence can still shift or drift over time—especially as a simulation runs longer". The Note "A world that follows" covers it.
Achieved whenan outside group publishes how often outcomes in Genie 4's scenes match real robot or machine runs on a fixed set of tasks.
Claimed by
Google DeepMind, October 2026
Open
01
Forecasts posted before the fact
A small physical rig, such as two water tanks with a pump and valves, and a twin of it that publishes each day a forecast of the next 24 hours of levels and flows, with uncertainty bands, before those hours happen. A script scores every forecast against what was measured.
The briefCredit · unclaimed02
Phone-built copies on the bench
For ten tabletop robot tasks, build a simulation of each from a phone video, run the same robot program 20 times in the copy and 20 times on the real arm, and publish both success rates side by side.
The briefCredit · unclaimed
Open 01 · the brief
Forecasts posted before the fact
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A small physical rig, such as two water tanks with a pump and valves, and a twin of it that publishes each day a forecast of the next 24 hours of levels and flows, with uncertainty bands, before those hours happen. A script scores every forecast against what was measured.
Why it is still open
The National Academies report found "a lack of standards in reporting VVUQ as well as a lack of consideration of confidence in modeling outputs", and recommends that advances be "evaluated using real-world use cases and testbeds".
Built with what exists
OpenModelica for the physics model, exported as an FMU and run with FMPy; a Raspberry Pi with level and flow sensors; Mosquitto, an open-source MQTT broker; InfluxDB and Grafana to store and plot the data; scikit-learn or PyTorch for a data-driven correction to the physics.
First step
Build the two-tank rig and run the Modelica model against one day of logged levels.
Done looks like
90 days of forecasts, each committed to a public Git repository before its day begins, with the mean error and the share of measured values that fell inside the 90% bands (between 85% and 95% counts as honest bands) published.
Open 02 · the brief
Phone-built copies on the bench
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For ten tabletop robot tasks, build a simulation of each from a phone video, run the same robot program 20 times in the copy and 20 times on the real arm, and publish both success rates side by side.
Why it is still open
RoboSimStudio's page gives no measured comparison of its copies with the real scenes. In the VABAR benchmark paper of 4 November 2026, from UC Berkeley, NVIDIA and Siemens, one robot program written by a person succeeded 87.0% of the time on the Newton simulator and 52.5% on Isaac Lab-Arena for the same task (Syringe Packing, level 1, Table II); the authors note that the two versions differ in "texture, lighting, asset and robot implementations".
Built with what exists
RoboSimStudio (Apache 2.0); the Newton physics engine and NVIDIA Isaac Lab; a low-cost arm such as the SO-101 run with Hugging Face's LeRobot (see the Note "No download"); a phone camera; Claude Code or Codex.
First step
Film one task, such as stacking two cups, and build its copy with RoboSimStudio.
Done looks like
For ten tasks, success rates from 20 runs in the copy and 20 on the real arm, the measured gap between object positions in the copy and in the video, and all code, videos and copies published.