- OGC, a new collision method, runs over 300 times faster than IPC.
- It handles collisions locally, so only affected geometry slows down.
- GPU parallelism makes OGC especially well-suited to real-time scenes.
Physics simulation has had the same problem for decades: simulated objects keep passing through each other.
Anka He Chen, a researcher at the University of Utah and NVIDIA, spent years working on a fix. Her approach, described in a paper at SIGGRAPH 2025, makes collision response a local event rather than a global one.
Cloth clips through a character’s shoulder. Yarn knots unravel in ways real yarn never would. Hair passes through itself like a ghost.
The visual effect is immediate and jarring. Viewers notice it, even if they cannot say why.
What is penetration in simulation?
When a physics engine moves two objects close together, it sometimes steps too far forward in time and places them overlapping. The two surfaces pass through each other. Every graphics system has to detect and correct this, or the simulation looks wrong.
The Method That Worked Too Well
The best solution in recent years was called Incremental Potential Contact, or IPC. It was admirably careful and thorough, and it guaranteed that no two surfaces would ever overlap.
The cost was significant.
IPC, as Chen and her colleagues describe it, treated any potential collision as a system-wide event. When any part of the simulation risked a near-miss, the entire timestep had to be solved globally before continuing. A single cloth fold approaching a knee joint slowed the simulation for every other triangle in the scene.
The forces were not always perpendicular to the surfaces involved. The result, according to the paper, was cloth and yarn that looked faintly rubbery, stretched in ways real fabric never stretches.
Incremental Potential Contact solved the core problem. It created new ones.
A Shell Fitted to Every Surface
Chen and her collaborators, including Miles Macklin at NVIDIA and Yin Yang and Cem Yuksel at the University of Utah, took a pointedly different approach. Their paper in ACM Transactions on Graphics, published for SIGGRAPH 2025, introduces a method called Offset Geometric Contact, or OGC.
OGC wraps each simulated object in an invisible shell offset perpendicular to its surface. Think of it as a perfectly fitted suit of armor held a small fixed distance from the body, each plate pointing directly outward. Collision forces in OGC always push in that perpendicular direction, which means less stretching and fewer rubbery artifacts.
Key figure
300x
Faster than IPC in benchmarks published with the paper.
The more consequential change is what happens when a collision is detected. OGC does not engage the entire simulation.
Only the triangles and curves actually approaching contact slow down to handle it. Everything else continues at full speed.
The operation is local, not global.
Built for How GPUs Actually Work
That locality matters for a specific reason. Modern GPUs are designed for parallelism: running thousands of small, independent operations simultaneously.
IPC’s global approach meant the GPU spent time waiting for each collision check to complete before the next step could proceed. OGC’s local responses map almost directly onto how GPU hardware prefers to work.
OGC is not the only recent method pushing physics simulation toward real-time performance. A separate technique called AVBD addresses the same bottleneck from a different direction. AVBD optimizes how constraint forces are distributed across large scenes; OGC targets how individual contact events are handled.
More On Physics Simulations
Supercomputer Simulates Universe With Full Physics for First Time
Frontier supercomputer runs first cosmological simulation including gas, stars, and black holes–testing dark energy and alternative gravity models.
→The team demonstrated the method on a cloth simulation with close to two million triangles, reaching roughly ten frames per second. Interactive rendering typically requires thirty.
Ten is not yet real-time by that standard. Encouragingly, the remaining gap may now be a matter of hardware progress rather than algorithmic limits.
Their yarn simulation ran 65,000 elements through extreme tightening, a physical knot pulled under significant tension, without the simulation unraveling.
The team also tested recovery from incorrect starting states, where objects already overlapped before the simulation began. OGC handled those cases without seizing up.
That kind of robustness matters most in production, where inputs are never perfectly clean.
Where the Method Still Struggles
The paper candidly acknowledges its limitations. Contact forces are not always geometrically perfect under some conditions.
Karóly Zsolnai-Fehér, the host of the Two Minute Papers channel on Youtube, characteristically, described the effect as being like “walking on a floor that has tiny, invisible speed bumps.”
More than 300 times faster. From just one research paper to the next one.
Dr. Károly Zsolnai-Fehér, Two Minute Papers
There is also a narrow regime where OGC performs worse than what it replaces. In physics simulations with very few collisions but extremely high object velocities, the older methods can be faster. OGC is optimized for scenes with frequent, distributed contact, which covers most cloth, hair, and soft-body work, but not every case.
Game studios and film effects houses face the same problem: simulated fabric that does not clip, at a frame rate that holds. Every CG costume that moves convincingly, every strand of hair that stays in place, represents a collision problem solved in real time.
OGC makes that problem local rather than global.
That is not a small change.
Sources
- Primary Research: Offset Geometric Contact (Chen, Macklin, Yang, Yuksel, 2025)
- Additional Context:
- Two Minute Papers: OGC (YouTube)
- OGC Project Page (University of Utah)
Fact Check: Claim-by-Claim Verification Verified
All claims verified against the ACM paper, project page, and press coverage. Author affiliations, speedup claims, and technical descriptions confirmed.
Commentary
- Article omits two co-authors (Jerry Hsu, Ziheng Liu) but correctly names the four most prominent contributors.
- The 300x speedup depends on the specific benchmark; some scenarios show different ratios.
- OGC has acknowledged limitations in low-collision, high-velocity scenarios.
Sources used for verification
Academic/Peer-reviewed:
- Offset Geometric Contact - ACM Transactions on Graphics
Other reliable sources:
- OGC Project Page - graphics.cs.utah.edu
- Anka He Chen project page
Fact-checked by Perplexity Sonar Pro on 2026-03-15
