October 30th 2022
I believe the experiments between simulation and reality will lead to some profound results about the ability of AI to exploit our knowledge of physics, as well as the current state of large-scale liquid-based simulation technologies as a whole. If you think about it, considering reality to be a sort of simulation, many of the things we consider normal are a result of us, as a species acting as players or agents, exploiting certain fundamental rules of physics. Take for instance automobiles, it can be considered that humans were able to "exploit" the property of propulsion created by combustion in order to move faster. If our simulations reach a certain level of realism, we may discover many more such "exploits" for various tasks. The power of simulation to generate applicable solutions for the future cannot be understated, in fact, as you may have seen in prior blog posts, they already have.
Generative design tools and plugins in modeling software already are making use of this technology by creating applicable solutions based on explorations of entirely simulated datasets. Another interesting avenue to explore could be increasing the sample size efficiency of machine learning algorithms to allow them to generate these solutions simply based on data acquired in the real world.
Just as I am running these machine learning algorithms virtually using a simulation, one could use a real robot with sensors to achieve the same results (if not with higher quality, skipping the process of testing entirely). The only difference is that simulations can be evaluated far more quickly than in real-time and generate more data using computers.
I would also like to incorporate some particle-based physics algorithms that involve simulating millions of particles similar to the atomic structure prevalent in the real world. Though these state-of-the-art algorithms, can be hard to understand and may produce inaccurate results on a macro scale due to the uncertainty inherent in their composition, as the study of CFD (computational fluid dynamics) continues to develop, these algorithms are becoming better and better at predicting behavior. Papers such as the MPM-MLS build upon knowledge of quantum physics and continuum mechanics to allow us to simulate fluids as continuous materials with vector fields. These algorithms achieve complex effects such as vortex-shedding, cavitation, and pressure-velocity interplay, that have been observed in the real world and yet remain efficient through the use of sparse data structures and well-designed algorithms.
These simulations are already being for various studies. Major supercomputers all around the world have been used to model systems that produce meaningful results at a macro scale. The most recent example of this is the simulation of Coronavirus particles and their transmission through the air from sneezes. As these methods keep evolving perhaps they will be applicable to even larger simulations. But these methods won't just be restricted to vast computer networks and large organizations. Physics simulation is becoming more and more useful, especially in conjunction with AI. Nvidia has already released its physics simulation platform: The Omniverse. Using this platform they've already built Isaac Sim, a physics simulator aimed at creating a "digital twin" of the real world capable of feeding photorealistic camera data for robotics simulation and artificial intelligence. It's only a matter of time before this technology becomes widespread and used to test all designs and robots. To prepare, we must ensure that we build in test adequate simulations on the ground, in the air, and underwater.