Physics of Video Games
A physics and simulation research lab for talented middle and high schoolers who play video games
In-person in Princeton, NJ · or hybrid from anywhere
The Opportunity
You play video games. In this lab you design studies to test the physics in the games you play, and you build simulations of your own that can become games.
Not every game simulates physics realistically. Many take shortcuts, and even the best engines make trade-offs to run smoothly at 60 frames per second.
- How fast something falls in a game is rarely what physics predicts. It depends on how the code steps time forward, and on whichever shortcut for air resistance was cheap enough to ship. Gravity is the most often violated principle in released software.
- How a pile of crates settles depends on how many times the code works out what is touching what before the frame is due. Two engines given the same scene disagree.
- A sea surface is waves of many sizes crossing each other, and what the water does below matters too. Real-time models have only begun to compute that.
- A storm cloud grows because warm air rises, because droplets and ice form inside it, and because those particles pick up charge. That charge puts the lightning in it.
- Sand pouring, snow packing, glass shattering: film studios compute those offline, taking as long as they need. A game has about sixteen milliseconds. Closing that gap is unfinished.
- Neither relativity nor quantum mechanics shows up at the scale a person lives at. Anyone who meets them meets a representation somebody built, and how well people reason inside one is open.
These are just examples; the field is wider than this.
Physics of Video Games is one of the SoTS Research Labs, which follow the same methodology and format: see Princeton Labs for how semesters work and expectations from students participating in the program.
Recent Work in This Field
This is an active field. Recent work reveals what released engines get wrong and how often, how researchers simulate oceans and storm clouds, what it costs to run these simulations inside a single frame, and simulations built specifically to be played.
Recent publications
What the engines get wrong
- A taxonomy of physics failures in software built on physics engines found gravity violations in 40 percent of cases and violations of Newton's laws in 28.2 percent, across 17 categories (Li et al., 2025).
- A benchmark of 880 gameplay clips with glitches spans four domains, from mechanics to material properties, covering 12 kinds of physical commonsense (Cao et al., 2024).
- Replaying the same initial state, action trace, and camera path in 12 scenes under different gravity settings produced 60 million frames where only gravity had changed (Hu et al., 2026).
- A review of nine simulation frameworks rated Unity as easy to use but limited in scalability and simulation fidelity. It also noted that many published comparisons come from the engines' own developers (Kaup et al., 2024).
What is hard to simulate
- A real-time ocean framework couples solid bodies to the free surface and includes the velocity of water at any depth — something earlier models left out (Algis et al., 2025).
- The first physically based simulation of storm systems for graphics combines cloud microphysics with hydrometeor electrification. This means the lightning emerges naturally from the storm (Hao et al., 2025).
- A position-based method with smoothing kernels simulated von Mises plasticine, Drucker-Prager sand, Cam Clay snow, and brittle fracture, all in real time (Yu et al., 2024).
What it costs to run inside a frame
- A material point method pipeline built for modern GPUs handled 1.33 million particles in real time, at 60 frames a second or better (Fei et al., 2021).
- A simulator that learned from single-timestep predictions generalized to new initial conditions, thousands of timesteps, and ten times more particles than it trained on (Sanchez-Gonzalez et al., 2020).
- A hybrid of a numerical solver and a neural network ran interactive fluid simulations at real-time frame rates, steered by freehand sketches (Xu et al., 2025).
Simulations built to be played
- In a citizen science game on quantum optimal control, solutions that players optimized on their own devices performed roughly on par with the best standard optimizers running on a cluster (Jensen et al., 2021).
- An authoring tool turned static textbook diagrams into interactive simulations of optics, circuits, and kinematics, based on an elicitation study with seven physics instructors (Gunturu et al., 2024).
- A pedagogical treatment shows how to implement Lorentz transformations in relativistic visualization, using geometric algebra in 3+1 dimensions (Hamilton, 2021).
The Engines and the Tools
Here, you answer questions by measuring a released game from the outside, or by building the simulation and making it playable. Each approach checks the other. Measuring gives you a number to set against the equation it was meant to follow. A simulation you built can be measured the same way, which is how you find out whether it does what you think it does.
Finding the tool that fits your question is part of the work, so these are examples rather than a complete list:
- Measuring a video game from outside it: Tracker
- Published records of what engines get wrong: PhysGame and GameBench
- Writing the simulation: Taichi, NVIDIA Warp
- Turning it into something playable: Godot, Bullet, Box2D
Prerequisites
Open to high school, middle school, and home school students who play video games. No prior research experience or physics background is required.
Full requirements
- Ready to read real research papers
- Ready to design your own study, not follow pre-made assignments
- Ready to work with data and code on a computer
- Commitment to weekly meetings and work between sessions
- A computer that can run the video game or simulation for your project