arxiv: 2604.23520 · v1 · submitted 2026-04-26 · 💻 cs.GR

Recognition: unknown

Rethinking Collision Detection on GPU Ray Tracing Architecture

Artem Pelenitsyn, Durga Keerthi Mandarapu, Gilbert Bernstein, Isaac Fuksman, Milind Kulkarni

Pith reviewed 2026-05-08 05:01 UTC · model grok-4.3

classification 💻 cs.GR

keywords collision detectionray tracingGPUbounding volume hierarchyparticle simulationdiscrete collision detectionproxy spheres

0 comments

The pith

Mochi uses per-object proxy spheres to decouple bounding volumes from collision radius, enabling efficient detection for non-uniform spheres on GPU ray tracing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Mochi, a new method to perform discrete collision detection for spherical particles with varying radii on modern GPU ray tracing hardware. Previous approaches reduced the problem to fixed-radius neighbor search, but this fails for non-uniform particles, forcing large bounding boxes and poor performance. Mochi exploits the symmetry of collision relations and introduces proxy spheres to allow tighter bounding volume hierarchies without missing any true collisions. It is proven sound and integrated into a particle simulation pipeline, showing speedups over existing methods. This matters because efficient collision detection is key to scalable simulations in physics, graphics, and related fields.

Core claim

Mochi reformulates discrete collision detection as a reduction on ray tracing architecture by using per-object proxy spheres that decouple the BVH bounding volumes from the collision search radius. This supports both uniform and non-uniform spherical particles efficiently, is provably sound, and guarantees detection of all true collisions while enabling tighter bounding boxes.

What carries the argument

per-object proxy spheres that decouple BVH bounding volumes from the collision search radius, exploiting symmetry of collision relations

If this is right

Consistent speedups over state-of-the-art BVH-based and RT-based DCD implementations in large-scale particle workloads.
Generalization of prior RT-based neighbor search formulations without their limitations for non-uniform spheres.
Provable soundness ensuring no true collisions are missed.
Integration into end-to-end particle simulation pipelines with improved performance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

May enable scaling to larger simulations in molecular dynamics or granular materials where radius variation is common.
Could inspire similar proxy-based techniques for other hardware-accelerated traversal problems beyond collisions.
The symmetry exploitation might apply to other symmetric relations in computational geometry.

Load-bearing premise

That per-object proxy spheres can always be chosen to produce tighter BVHs while preserving the guarantee that every true collision is reported, without introducing unacceptable overhead or false negatives for arbitrary radius distributions.

What would settle it

Finding a configuration of particles with varying radii where Mochi either misses a true collision or performs worse than previous RT-based methods on the same hardware.

Figures

Figures reproduced from arXiv: 2604.23520 by Artem Pelenitsyn, Durga Keerthi Mandarapu, Gilbert Bernstein, Isaac Fuksman, Milind Kulkarni.

**Figure 1.** Figure 1: Collisions occur within a distance of 2𝑟 and 𝑟𝑖 + 𝑟𝑚𝑎𝑥 from sphere centers for uniform (left) and non-uniform (right) spheres, respectively. false positives while preserving correctness, improving traversal efficiency without missing collisions. 3 Collision Detection using Proxy Spheres This section describes how Mochi performs collision detection for spherical particles on RT architecture using an interme… view at source ↗

**Figure 3.** Figure 3: Particles under freefall after eight seconds: using view at source ↗

**Figure 5.** Figure 5: Performance comparison of Mochi DCD kernel view at source ↗

**Figure 6.** Figure 6: Performance comparison of Mochi DCD kernel view at source ↗

**Figure 7.** Figure 7: Performance comparison of Mochi against Taichi view at source ↗

**Figure 9.** Figure 9: Performance comparison of Mochi against Taichi view at source ↗

**Figure 8.** Figure 8: Performance comparison of Mochi against Taichi view at source ↗

**Figure 10.** Figure 10: Performance comparison of Mochi against Taichi view at source ↗

**Figure 11.** Figure 11: Performance comparison of Mochi against Taichi view at source ↗

**Figure 12.** Figure 12: Build and DCD times of Mochi as the number of view at source ↗

read the original abstract

Discrete Collision Detection (DCD) is a fundamental task in several domains including particle-based physics simulations. Efficient DCD uses indexing structures such as Bounding Volume Hierarchy (BVH), but accelerating irregular BVH traversals demands meticulous efforts to achieve performance. Modern GPUs feature Ray Tracing (RT) architecture that provides hardware acceleration for BVH traversal and optimized drivers for BVH construction. Recent work has attempted to exploit RT architecture to accelerate DCD on spherical particles by reducing DCD to fixed-radius neighbor search. However, this reduction breaks down for particles with different radii, necessitating the use of large bounding boxes that result in a higher number of duplicate collisions and poor performance. To address these limitations, we present Mochi, a new reduction that reformulates DCD on RT architecture by exploiting the symmetry of collision relations to support both uniform and non-uniform spherical particles efficiently. Mochi introduces per-object proxy spheres that decouple BVH bounding volumes from the collision search radius, enabling significantly tighter bounding boxes without sacrificing correctness. Mochi is provably sound and guarantees that all true collisions are detected. We integrate Mochi into an end-to-end particle simulation pipeline and evaluate it across large-scale particle workloads, showing consistent speedups over state-of-the-art BVH-based and RT-based DCD implementations. Mochi generalizes prior RT-based neighbor search formulations while avoiding their fundamental limitations for non-uniform spheres.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Mochi's per-object proxy spheres give a workable way to run variable-radius collision detection on RT hardware without oversized bounds, but the tightness guarantee for wide radius spreads needs the full construction details.

read the letter

The paper's main advance is a reduction that exploits collision symmetry and adds per-object proxy spheres so the BVH built for ray-tracing traversal can stay tighter than the usual max-radius envelope while still reporting every true pair. This directly fixes the problem the abstract describes in prior fixed-radius RT neighbor search when particle sizes differ. They integrate the method into a full particle simulation pipeline and report consistent speedups against both classic BVH and earlier RT approaches on large workloads. That practical end-to-end evaluation is the part that matters most for people who actually run these simulations. The formulation looks like a clean generalization rather than a narrow special case. The soundness claim is stated outright, which is better than many performance papers that skip it. On the downside, the abstract gives no proof sketch or error analysis, so the central guarantee—that proxy radii can always be chosen to produce strictly tighter nodes without false negatives—rests on whatever construction they use in the body. The stress-test worry about wide radius variation forcing proxies back toward conservative sizes is real and should be checked against their worst-case examples. If the proxy formula handles arbitrary distributions without extra overhead or missed pairs, the speedups should hold; otherwise the benefit shrinks. This is aimed at graphics and simulation researchers who care about GPU ray-tracing primitives for neighbor search or collision. A reader already working with RT cores or particle codes will find the reduction and the reported numbers useful. It is solid enough on the algorithmic side and the evaluation to deserve referee time rather than a desk reject.

Referee Report

2 major / 1 minor

Summary. The paper presents Mochi, a reformulation of discrete collision detection (DCD) for spherical particles on GPU ray-tracing hardware. It exploits symmetry in collision relations and introduces per-object proxy spheres to decouple BVH bounding volumes from the collision search radius, enabling tighter bounding boxes for both uniform and non-uniform radii. The method is claimed to be provably sound with a guarantee that all true collisions are detected, generalizes prior fixed-radius RT-based neighbor search, and is integrated into an end-to-end particle simulation pipeline with reported consistent speedups over state-of-the-art BVH and RT implementations.

Significance. If the soundness guarantee and performance claims hold, the work would meaningfully advance hardware-accelerated collision detection for irregular particle systems by removing the need for conservative max-radius bounds. The explicit integration into a full simulation pipeline and evaluation on large-scale workloads provide practical grounding that strengthens the contribution beyond a purely algorithmic reformulation.

major comments (2)

[Abstract] Abstract: The central claim that 'Mochi is provably sound and guarantees that all true collisions are detected' is load-bearing for the entire contribution, yet the manuscript provides no theorem statement, proof sketch, or error analysis to substantiate that per-object proxy spheres preserve detection of every pair satisfying |p_i - p_j| ≤ r_i + r_j while producing strictly tighter BVHs.
[Abstract] Abstract and method description: For arbitrary radius distributions, the proxy-sphere construction must ensure that a small-radius particle's proxy still intersects rays or volumes from arbitrarily large partners without false negatives or reverting to max-radius conservatism. The symmetry exploitation is noted but no explicit worst-case bounding formula or counter-example analysis is supplied to confirm tightness is always achieved.

minor comments (1)

[Abstract] Abstract: The evaluation is summarized as 'consistent speedups' and 'large-scale particle workloads' without naming particle counts, radius variance ranges, hardware platform, or baseline implementations, which obscures assessment of the claimed gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and will revise the manuscript to include the requested formal elements supporting the soundness claims.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that 'Mochi is provably sound and guarantees that all true collisions are detected' is load-bearing for the entire contribution, yet the manuscript provides no theorem statement, proof sketch, or error analysis to substantiate that per-object proxy spheres preserve detection of every pair satisfying |p_i - p_j| ≤ r_i + r_j while producing strictly tighter BVHs.

Authors: We agree that the current manuscript lacks an explicit theorem statement, proof sketch, and error analysis for the soundness guarantee. In the revised version we will add a dedicated subsection (or appendix) containing a formal theorem: for any particles i and j, if |p_i - p_j| ≤ r_i + r_j then the proxy-sphere construction ensures the ray-tracing traversal detects the pair. The proof sketch will derive this from the symmetry of the collision relation and the per-object proxy radius definition that decouples BVH bounds from search radii. A short error analysis confirming absence of false negatives will also be included. revision: yes
Referee: [Abstract] Abstract and method description: For arbitrary radius distributions, the proxy-sphere construction must ensure that a small-radius particle's proxy still intersects rays or volumes from arbitrarily large partners without false negatives or reverting to max-radius conservatism. The symmetry exploitation is noted but no explicit worst-case bounding formula or counter-example analysis is supplied to confirm tightness is always achieved.

Authors: We will expand the method description to supply the explicit worst-case bounding formula for proxy radii under arbitrary radius distributions. The formula will show that each particle's proxy is sized to guarantee intersection with any valid partner (including extreme radius ratios) while remaining strictly tighter than a global max-radius bound. We will also add a short worst-case analysis together with a brief counter-example verification demonstrating that the symmetry-based construction preserves completeness without false negatives. revision: yes

Circularity Check

0 steps flagged

No circularity: Mochi's proxy-sphere reformulation is an independent algorithmic construction

full rationale

The paper's central contribution is a new reduction of discrete collision detection to RT-accelerated BVH traversal that exploits collision symmetry via per-object proxy spheres. This construction is presented as a direct algorithmic reformulation that decouples bounding volumes from search radii while preserving soundness; no equations, parameters, or uniqueness claims are shown to reduce to fitted inputs, self-definitions, or load-bearing self-citations. The provable-soundness guarantee is asserted from the proxy definition itself rather than from any prior result by the same authors, and the evaluation compares against external baselines without renaming known patterns or smuggling ansatzes. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The approach rests on standard assumptions about collision symmetry and BVH/RT hardware behavior plus one new invented construct.

axioms (2)

domain assumption Collision relations are symmetric: if particle A collides with B then B collides with A.
Invoked to reformulate DCD as a neighbor search on RT hardware.
domain assumption Hardware BVH traversal on modern GPUs is efficient for fixed-radius queries when bounding volumes are tight.
Basis for claiming performance gains from tighter boxes.

invented entities (1)

per-object proxy spheres no independent evidence
purpose: Decouple BVH bounding volumes from the actual collision search radius to allow tighter boxes for non-uniform particles
New construct introduced by the paper; no independent evidence provided beyond the claim of soundness.

pith-pipeline@v0.9.0 · 5560 in / 1340 out tokens · 38088 ms · 2026-05-08T05:01:51.750841+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

48 extracted references · 9 canonical work pages

[1]

Ciprian Apetrei. 2014. Fast and Simple Agglomerative LBVH Construction. In Computer Graphics and Visual Computing (CGVC), Rita Borgo and Wen Tang (Eds.). The Eurographics Association. doi:10.2312/cgvc.20141206

work page doi:10.2312/cgvc.20141206 2014
[2]

M Breuer and N Almohammed. 2015. Modeling and simulation of particle agglom- eration in turbulent flows using a hard-sphere model with deterministic collision detection and enhanced structure models.International Journal of Multiphase Flow73 (2015), 171–206

2015
[3]

Tyson Brochu, Essex Edwards, and Robert Bridson. 2012. Efficient geometrically exact continuous collision detection.ACM Transactions on Graphics (TOG)31, 4 (2012), 1–7

2012
[4]

Floyd M Chitalu, Christophe Dubach, and Taku Komura. 2020. Binary Ostensibly- Implicit Trees for Fast Collision Detection. InComputer Graphics Forum, Vol. 39. Wiley Online Library, 509–521

2020
[5]

Peter A Cundall and Otto DL Strack. 1979. A discrete numerical model for granular assemblies.geotechnique29, 1 (1979), 47–65

1979
[6]

David Eberly. 2008. Intersection of Ellipsoids. https://www.geometrictools.com/ Documentation/IntersectionOfEllipsoids.pdf

2008
[7]

Evangelou, G

I. Evangelou, G. Papaioannou, K. Vardis, and A. A. Vasilakis. 2021. Fast Radius Search Exploiting Ray Tracing Frameworks.Journal of Computer Graphics Tech- niques (JCGT)10, 1 (5 February 2021), 25–48. http://jcgt.org/published/0010/01/ 02/

2021
[8]

Liang Geng, Rubao Lee, and Xiaodong Zhang. 2024. RayJoin: Fast and Precise Spatial Join. InProceedings of the 38th ACM International Conference on Super- computing. 124–136

2024
[9]

Liang Geng, Rubao Lee, and Xiaodong Zhang. 2025. LibRTS: A Spatial Index- ing Library by Ray Tracing. InProceedings of the 30th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming. 396–411

2025
[10]

Justus Henneberg and Felix Schuhknecht. 2023. RTIndeX: Exploiting Hardware- Accelerated GPU Raytracing for Database Indexing.Proceedings of the VLDB Endowment16, 13 (2023), 4268–4281

2023
[11]

Everton Hermann, François Faure, and Bruno Raffin. 2008. Ray-traced collision detection for deformable bodies. InGRAPP 2008-3rd International Conference on Computer Graphics Theory and Applications. INSTICC, 293–299

2008
[12]

HIP-RT. 2024. Tutorial 16, fluid simulation on HIP RT SDK. https://github.com/ GPUOpen-LibrariesAndSDKs/HIPRTSDK

2024
[13]

Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. 2019. Taichi: a language for high-performance computation on spatially sparse data structures.ACM Transactions on Graphics (TOG)38, 6 (2019), 201

2019
[14]

Markus Ihmsen, Jens Orthmann, Barbara Solenthaler, Andreas Kolb, and Matthias Teschner. 2014. SPH fluids in computer graphics. (2014)

2014
[15]

Youngjun Kim, Sang Ok Koo, Deukhee Lee, Laehyun Kim, and Sehyung Park
[16]

In2010 International Conference on Cyberworlds

Mesh-to-mesh collision detection by ray tracing for medical simulation with deformable bodies. In2010 International Conference on Cyberworlds. IEEE, 60–66
[17]

Homogeneous middles vs

C. Lauterbach, M. Garland, S. Sengupta, D. Luebke, and D. Manocha. 2009. Fast BVH Construction on GPUs.Computer Graphics Forum(2009). doi:10.1111/j.1467- 8659.2009.01377.x

work page doi:10.1111/j.1467- 2009
[18]

Christian Lauterbach, Qi Mo, and Dinesh Manocha. 2010. gProximity: hierarchical GPU-based operations for collision and distance queries. InComputer Graphics Forum, Vol. 29. Wiley Online Library, 419–428

2010
[19]

François Lehericey, Valérie Gouranton, and Bruno Arnaldi. 2013. New iterative ray-traced collision detection algorithm for gpu architectures. InProceedings of the 19th ACM Symposium on Virtual Reality Software and Technology. 215–218

2013
[20]

François Lehericey, Valérie Gouranton, and Bruno Arnaldi. 2013. Ray-traced collision detection: Interpenetration control and multi-gpu performance. In5th Joint Virtual Reality Conference of EuroVR-EGVE. 1–8

2013
[21]

François Lehericey, Valérie Gouranton, and Bruno Arnaldi. 2015. GPU Ray-Traced Collision Detection: Fine Pipeline Reorganization. InProceedings of 10th Interna- tional Conference on Computer Graphics Theory and Applications (GRAPP’15)

2015
[22]

Fuchang Liu, Takahiro Harada, Youngeun Lee, and Young J Kim. 2010. Real- time collision culling of a million bodies on graphics processing units.ACM Transactions on Graphics (TOG)29, 6 (2010), 1–8

2010
[23]

Yangming Lv, Kai Zhang, Ziming Wang, Xiaodong Zhang, Rubao Lee, Zhenying He, Yinan Jing, and X Sean Wang. 2024. RTScan: Efficient Scan with Ray Tracing Cores.Proceedings of the VLDB Endowment17, 6 (2024), 1460–1472

2024
[24]

Durga Keerthi Mandarapu, Vani Nagarajan, Artem Pelenitsyn, and Milind Kulka- rni. 2024. Arkade: k-Nearest Neighbor Search With Non-Euclidean Distances using GPU Ray Tracing. InProceedings of the 38th ACM International Conference on Supercomputing. 14–25

2024
[25]

Daniel Meister, Shinji Ogaki, Carsten Benthin, Michael J Doyle, Michael Guthe, and Jiří Bittner. 2021. A survey on bounding volume hierarchies for ray tracing. InComputer Graphics Forum, Vol. 40. Wiley Online Library, 683–712

2021
[26]

Vani Nagarajan, Rohan Gangaraju, Kirshanthan Sundararajah, Artem Pelenitsyn, and Milind Kulkarni. 2025. RT-BarnesHut: Accelerating Barnes-Hut Using Ray- Tracing Hardware. InProceedings of the 30th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming. 43–56

2025
[27]

Vani Nagarajan and Milind Kulkarni. 2023. RT-DBSCAN: Accelerating DBSCAN using Ray Tracing Hardware.CoRRabs/2303.09655 (2023). arXiv:2303.09655 doi:10.48550/arXiv.2303.09655

work page doi:10.48550/arxiv.2303.09655 2023
[28]

Vani Nagarajan, Durga Mandarapu, and Milind Kulkarni. 2023. RT-kNNS Un- bound: Using RT Cores to Accelerate Unrestricted Neighbor Search. InProceedings of the 37th International Conference on Supercomputing, ICS 2023, Orlando, FL, USA, June 21-23, 2023, Kyle A. Gallivan, Efstratios Gallopoulos, Dimitrios S. Nikolopou- los, and Ramón Beivide (Eds.). ACM, 2...

work page doi:10.1145/3577193.3593738 2023
[29]

Steven G. Parker, Heiko Friedrich, David Luebke, Keith Morley, James Bigler, Jared Hoberock, David McAllister, Austin Robison, Andreas Dietrich, Greg Humphreys, Morgan McGuire, and Martin Stich. 2013. GPU ray tracing.Commun. ACM56, 5 (may 2013), 93–101. doi:10.1145/2447976.2447997

work page doi:10.1145/2447976.2447997 2013
[30]

RayTracedSPH. 2024. Real-time fluid simulation using smoothed particle hydro- dynamics. https://github.com/StarsX/RayTracedSPH

2024
[31]

Xin Shi and Li Tian. 2022. Towards robotic assembly: collision detection between each part of the parallel groove clamp.The International Journal of Advanced Manufacturing Technology119, 7 (2022), 4349–4358

2022
[32]

Xuri Shi, Kai Zhang, X Sean Wang, Xiaodong Zhang, and Rubao Lee. 2024. RTCUDB: Building Databases with RT Processors.arXiv preprint arXiv:2412.09337 (2024)

work page arXiv 2024
[33]

Sizhe Sui, Luis Sentis, and Andrew Bylard. 2025. Hardware-accelerated ray tracing for discrete and continuous collision detection on gpus. In2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 16133–16139

2025
[34]

Tzvetomir Ivanov Vassilev. 2021. Real-time velocity based cloth simulation with ray-tracing collision detection on the graphics processor. In2021 International Conference on Information Technologies (InfoTech). IEEE, 1–5

2021
[35]

Monan Wang and Jiaqi Cao. 2021. A review of collision detection for deformable objects.Computer Animation and Virtual Worlds32, 5 (2021), e1987

2021
[36]

Tongtong Wang, Zhihua Liu, Min Tang, Ruofeng Tong, and Dinesh Manocha
[37]

InComputer Graphics Forum, Vol

Efficient and Reliable Self-Collision Culling Using Unprojected Normal Cones. InComputer Graphics Forum, Vol. 36. Wiley Online Library, 487–498
[38]

Xinlei Wang, Min Tang, Dinesh Manocha, and Ruofeng Tong. 2018. Efficient BVH- based collision detection scheme with ordering and restructuring. InComputer graphics forum, Vol. 37. Wiley Online Library, 227–237

2018
[39]

Sean Wang

Ziming Wang, Kai Zhang, Yangming Lv, Yinglong Wang, Zhigang Zhao, Zhenying He, Yinan Jing, and X. Sean Wang. 2024. RTOD: Efficient Outlier Detection With Ray Tracing Cores.IEEE Transactions on Knowledge and Data Engineering36, 12 (2024), 9192–9204. doi:10.1109/TKDE.2024.3453901

work page doi:10.1109/tkde.2024.3453901 2024
[40]

2013.New geometric data structures for collision detection and haptics

René Weller. 2013.New geometric data structures for collision detection and haptics. Springer Science & Business Media

2013
[41]

René Weller, Nicole Debowski, and Gabriel Zachmann. 2017. kDet: Parallel constant time collision detection for polygonal objects. InComputer Graphics Forum, Vol. 36. Wiley Online Library, 131–141

2017
[42]

Tsz Ho Wong, Geoff Leach, and Fabio Zambetta. 2014. An adaptive octree grid for GPU-based collision detection of deformable objects.The Visual Computer 30, 6 (2014), 729–738

2014
[43]

Anna Yershova and Steven M LaValle. 2007. Improving motion-planning algo- rithms by efficient nearest-neighbor searching.IEEE Transactions on Robotics23, 1 (2007), 151–157

2007
[44]

Stefan Zellmann, Martin Weier, and Ingo Wald. 2020. Accelerating Force-Directed Graph Drawing with RT Cores. In2020 IEEE Visualization Conference (VIS). 96–

2020
[45]

doi:10.1109/VIS47514.2020.00026

work page doi:10.1109/vis47514.2020.00026 2020
[46]

Shiwei Zhao, Zhengshou Lai, and Jidong Zhao. 2023. Leveraging ray tracing cores for particle-based simulations on GPUs.Internat. J. Numer. Methods Engrg. 124, 3 (2023), 696–713

2023
[47]

Shiwei Zhao and Jidong Zhao. 2023. Revolutionizing granular matter simulations by high-performance ray tracing discrete element method for arbitrarily-shaped particles.Computer Methods in Applied Mechanics and Engineering416 (2023), 116370

2023
[48]

Yuhao Zhu. 2022. RTNN: Accelerating Neighbor Search Using Hardware Ray Trac- ing. InProceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming(Seoul, Republic of Korea)(PPoPP ’22). Association for Computing Machinery, New York, NY, USA, 76–89. doi:10.1145/3503221.3508409

work page doi:10.1145/3503221.3508409 2022