The real bottleneck in physical AI
A robot can succeed many times in simulation and still fail the first time it reaches into the real world. This gap between simulated competence and real-world performance is the defining obstacle in physical AI today. Aicadium, a Singapore-based industrial AI company, has tested these systems first-hand. The pattern is consistent. The hard part is not imagining the world, but matching it.
The technical name for this obstacle is the sim-to-real gap. It arises because a simulation never perfectly reproduces reality. Lighting, friction, mass, and material behaviour all differ slightly, and small differences compound. Beneath the sim-to-real gap sits a deeper and more stubborn issue: the contact problem.
Why contact is so hard to model
Contact is hard because it is mathematically discontinuous. When an object strikes a surface, forces spike and velocities can reverse within a very short time. A simulator must approximate that sudden change, and every method of approximation involves a trade-off.
Hard contact, formulated as a precise constraint, is physically accurate but can become numerically unstable. Soft contact, which uses a penalty method, is more stable but allows objects to overlap slightly. Friction adds its own difficulty because it does not vary smoothly. Fast-moving objects can even pass through each other if the simulation steps forward too coarsely. The root cause is unavoidable. Contact is discontinuous, and no numerical method handles discontinuity cleanly.
What our experiments showed
In our own testing, we saw early signs of how hard generalisation can be. We trained pick-and-place policies entirely in simulation, and tested them in simulation. We did not deploy on a real robot. The setup used a robotic arm with two cameras and a standard imitation-learning model that copies expert demonstrations. On a single task with enough demonstrations, the policy succeeded reliably. When we asked it to transfer to an object it had not been trained on, performance dropped sharply.
The failures were instructive rather than random. The most common were incorrect object localisation, the gripper opening too early, and configurations the model had never seen in training. In our setup, the model did not generalise well beyond what it was trained on. We should be careful not to over-read one result. We tested a single model, and other models may well do better. The tentative lesson is that broad competence does not come for free. It appears to need relevant data and careful evaluation. A system that performs well on one task may still struggle on a closely related one.
We saw a similar pattern when testing one generalist vision-language-action model. This is a system that maps an instruction and a camera view directly to robot actions. Clear, direct instructions worked well. Selective or negated instructions, such as moving one object while leaving another untouched, seemed less reliable. A stronger and more explicit instruction improved the result. This is only an early observation, but it suggests how a task is specified may matter as much as the model behind it. We plan to test this further across more models and settings.
How world models and physics engines address the gap
World models and physics engines attack the sim-to-real gap from opposite directions. Physics engines narrow it by improving the accuracy of contact and material simulation. World models narrow it by generating diverse, realistic experiences and by transferring the visual appearance of the simulation closer to reality.
Neither fully closes the gap today. Contact-rich tasks, where precise force and geometry decide success, remain the domain of physics engines. World models contribute most where visual variety and data volume matter. A practical programme uses both, with clear eyes about which tool is carrying which load.
What does the sim-to-real gap mean for a business?
The sim-to-real gap means that an impressive simulation demonstration is not proof of real-world readiness. Before committing to deployment, leaders should ask a sharper question. How does a system perform on unseen configurations and contact-heavy tasks, not only on the scenarios it was trained for?
Can the contact problem be solved?
The contact problem cannot be eliminated, because contact is fundamentally discontinuous, but it can be managed. Newer methods, including differentiable and learned physics, are steadily improving how simulators handle contact. Progress over the next eighteen months will shape which physical-AI applications become commercially viable.
Aicadium’s view is that the sim-to-real gap is the right lens for evaluating physical AI today. The organisations that treat it as the central question, rather than a footnote, will set realistic expectations and avoid expensive disappointments. Honest evaluation against unseen conditions is the surest guide to what is ready and what is not.
