- AVBD solves physics simulations that defeated all previous methods.
- One iteration of AVBD outperforms 100 iterations of older techniques.
- The method simulates millions of objects at 100 fps.
Chris Giles was staring at a pendulum that refused to work.
The physics simulation should have been simple. A heavy ball on a chain of fifty linked segments, swinging back and forth. Pendulums are textbook mechanics. But on screen, the chain stretched like chewing gum, the ball hanging far below where it belonged.
The mass ratio between ball and chain was 50,000 to one. It was the kind of extreme mismatch that real-world engineering simulations encounter constantly, from crashing vehicles to swinging construction cranes.
Key figure
50,000:1
Mass ratio between ball and chain that broke every previous simulation method
When Textbook Physics Breaks Your Software
Giles, a researcher at the University of Utah's Graphics Lab, was working with computer graphics professor Cem Yuksel on improving real-time physics simulation. Their starting point was Vertex Block Descent, a method Yuksel's lab published in 2024 that could model millions of interacting points at remarkable speed.
VBD was fast, stable, and impressively parallel.
But scenarios involving extreme differences in mass or stiffness exposed a blind spot.
A house of cards hit by rolling balls generated so much artificial friction the cards barely moved, no matter how many computational iterations the solver ran. Chain mail that should have caught a thrown ball let it punch straight through.
A set of blocks connected by springs sagged under their own weight when the correct solution was a perfectly straight line. More iterations did not help. The method simply could not resolve the competing forces.
These were not obscure test cases. They were the kinds of interactions that game studios and visual effects houses need working correctly every day.
How AVBD Adjusts Its Grip
The fix came from a mathematical tool called an augmented Lagrangian formulation. Giles, Yuksel, and colleague Elie Diaz used it to build Augmented Vertex Block Descent, or AVBD. The work was published in ACM Transactions on Graphics, the field's top journal, and presented at SIGGRAPH 2025 in Vancouver.
What is an augmented Lagrangian formulation?
In physics simulation, a Lagrangian tracks how a system's energy changes over time. The "augmented" version adds penalty terms that grow when constraints are violated, letting the solver tighten its grip on problem areas without destabilizing the whole system.
Rather than applying uniform stiffness from the start, AVBD gradually increases how strictly it enforces constraints wherever violations are worst. Think of a security guard adjusting their stance. One person leaning on the barrier gets a light hand. A crowd pushing hard gets a firm, planted response. The effort scales with the problem.
The results were striking.
In the spring-connected blocks test, AVBD at a single iteration produced a nearly perfect solution. The previous method still sagged visibly after a hundred iterations. The chain mail caught the ball. The pendulum swung correctly. The house of cards collapsed on contact, exactly as physics demands.
Fast Physics Simulation on Consumer Hardware
AVBD handles complex scenes with millions of interacting objects at 100 frames per second on a single consumer graphics card. The source code is open and freely available, with a browser-based demo that anyone can run. For a SIGGRAPH paper, that level of accessibility is notable.
For Yuksel, an associate professor at the University of Utah who also serves as a research scientist at Roblox, the collaboration reflects a practical need.
More On Computer Graphics
When the Simulation Freezes: A New Method for Physics That Runs at Full Speed
A research team has solved one of computer graphics’ oldest frustrations with physics simulations.
→Game engines and visual effects pipelines require solvers that produce correct results without consuming entire render budgets.
A method that gets close to the right answer in one iteration, rather than struggling after a hundred, changes the cost equation entirely.
AVBD was built through mathematical insight, not machine learning. The entire improvement over the original VBD method happened in a single year of focused research by three people.
In a field increasingly shaped by neural networks, the method is a quiet reminder that classical techniques still have room to surprise.
Sources
- Primary Source: Augmented Vertex Block Descent (University of Utah Graphics Lab)
- Additional Context:
Fact Check: Claim-by-Claim Verification Verified
The article accurately describes the AVBD method, its developers, improvements over VBD, and key demonstrations, matching details from the primary research project page.
Commentary
- Article simplifies technical details (e.g., augmented Lagrangian explanation) appropriately for popular science audience without errors.
- SIGGRAPH 2025 and ACM Transactions publication details could not be directly verified due to access limits, but consistent with project timeline and venue norms.
- Exaggerated phrasing like "stumped every previous technique" is dramatic but supported by comparisons on project page.
Sources used for verification
Academic/Peer-reviewed:
- Augmented Vertex Block Descent (AVBD) - cs.utah.edu
- Augmented Vertex Block Descent - acm.org
Other reliable sources:
- AVBD Project Page - cs.utah.edu
Fact-checked by Perplexity Sonar Pro on 2026-03-11
