- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper presents PNEO, a learned simulator for collision dynamics among 3D deformable objects. PNEO is trained only on two-object collisions and applies a shared interaction model to every active contact pair at inference. An object-local neural operator encodes particle and object mechanics, after which a pairwise head predicts positive contact-stiffness and damping coefficients. These coefficients parameterize a quadratic contact potential and a Rayleigh dissipation term; the pairwise scalar terms are summed over the dynamic contact graph, differentiated into forces, and integrated with semi-explicit Euler. On a synthetic spring-mass benchmark, PNEO generally outperforms several graph-based simulators and transfers without retraining to scenes containing up to 50 objects, unseen stiffness values, and different point-cloud resolutions.
Critical strengths
-
Well-motivated compositional inductive bias. Reusing a shared pairwise interaction law over a changing contact graph is a natural way to separate scene cardinality from local contact mechanics. The zero-shot object-count experiments provide substantial evidence that this inductive bias improves transfer beyond the two-object training setting.
-
Broad evaluation of simulated mechanical fidelity. The paper evaluates position, penetration, strain, and contact-aligned rollout errors, and tests object-count, stiffness, point-resolution, and initial-condition shifts. In particular, the transfer from two-object training to scenes with up to 50 objects is an interesting empirical result.
Critical weaknesses
-
The implementation does not justify the claimed conservative neural-energy interpretation. The predicted contact stiffness is conditioned on both positions and velocities, including the normal relative velocity, but the coefficient is frozen when differentiating the quadratic potential. Consequently, the implemented force is not the full gradient of the stated state-dependent scalar energy, and a velocity-dependent stiffness does not define a conventional conservative potential. Re-predicting detached coefficients at every substep can also change the effective energy without an explicit conservation law. The paper reports no energy-balance, momentum, or dissipation measurements that would validate the claimed conservative energy exchange. The method can still be viewed as a structured pairwise force model, but its central energy-based interpretation is not established.
-
The benchmark strongly matches the hand-specified contact law, while key controls and scalability measurements are missing. The simulator uses frictionless pairwise penalty contact with quadratic penetration and normal damping, which closely matches PNEO's prescribed potential and dissipation forms. It is therefore unclear how much of the large advantage comes from learning rather than from encoding the data generator's contact law. The paper does not compare with an otherwise identical direct pairwise force predictor, fixed or analytic contact coefficients, or the closely related energy-based NEXUS method. It also motivates the work through computational scalability but reports no runtime, memory, training cost, or scaling curve. The absence of multiple-seed uncertainty further weakens the empirical comparison.
Questions for authors
-
Since the predicted stiffness depends on velocity and is detached during differentiation, in what precise mathematical sense is the resulting contact force conservative? Can the authors provide total-energy and dissipation measurements over rollouts?
-
How does PNEO compare with an architecture-matched pairwise model that directly predicts forces, and with fixed or oracle coefficients from the analytic penalty contact law? This comparison is necessary to isolate the value of the neural-energy parameterization.
-
What are the wall-clock runtime, peak memory, parameter count, and scaling behavior as the object count and point-cloud resolution increase? How does this compare with the reported baselines and NEXUS?
4 comments
W2 (1). Benchmark specificity and the prescribed contact law
To address the concern that our results are specific to the prescribed contact law in our deformable benchmark, we additionally evaluate a rigid-body adaptation of PNEO on MOVi-B, the benchmark used by RigidFormer. We retain the pairwise contact-graph formulation and contact-module design: predicting contact coefficients, constructing pairwise energies, and differentiating them to obtain forces for rigid-body velocity and pose updates. The table below reports position RMSE in meters (lower is better).
| Step size | Frame | RigidFormer | PNEO-rigid |
|---|---|---|---|
| 1 | 100 | 0.1079 | 0.0426 |
| 1 | 200 | 0.4357 | 0.2491 |
| 1 | 300 | 0.9411 | 0.6082 |
| 1 | 400 | 1.4061 | 0.9545 |
| 5 | 100 | 0.090 | 0.0378 |
| 5 | 200 | 0.4068 | 0.2411 |
| 5 | 300 | 0.8875 | 0.6016 |
| 5 | 400 | 1.3471 | 0.9506 |
| 10 | 100 | 0.0709 | 0.0309 |
| 10 | 200 | 0.3743 | 0.2289 |
| 10 | 300 | 0.8293 | 0.5838 |
| 10 | 400 | 1.2632 | 0.9261 |
We evaluate both methods at rollout horizons of 100, 200, 300, and 400 frames, using step sizes of 1, 5, and 10. PNEO-rigid achieves lower position RMSE than RigidFormer at every tested horizon and step size. We also train RigidFormer on our two-object deformable dataset, where it obtains RE@24/48 of 0.48/1.19, compared with 0.0633/0.1892 for our PNEO.
W2 (2) & Q2. Direct-force, analytic oracle and NEXUS comparisons
We compare PNEO with direct force prediction and an analytic oracle. The oracle uses the same graph-construction procedure as PNEO at test time, then assigns each structural edge an elastic coefficient using the dataset-generation rule, based on the given object stiffness and the edge's rest length. Each cell reports RE@24 / RE@48.
| Test M | Energy | Direct force | Analytic oracle |
|---|---|---|---|
| 2 | 0.0634 / 0.1824 | 0.0684 / 0.1797 | 0.0691 / 0.1885 |
| 3 | 0.0791 / 0.2050 | 0.0830 / 0.2051 | 0.0761 / 0.2197 |
| 5 | 0.0897 / 0.2233 | 0.0953 / 0.2297 | 0.0800 / 0.2142 |
PNEO achieves rollout errors comparable to this analytic oracle. Compared with direct force prediction, it achieves lower RE@24 at every tested object count () and lower RE@48 at ; direct force is slightly better at RE@48.
Comparison with NEXUS. We additionally compare PNEO with the closely related energy-based NEXUS method, both at its training object counts and on larger, unseen scenes. Tables R3 and R4 report RE@24 / RE@48:
Table R3. Accuracy against NEXUS trained at the corresponding object count.
| Test M | NEXUS training M | NEXUS ↓ | PNEO ↓ (train M=2) |
|---|---|---|---|
| 2 | 2 | 0.2124 / 0.4335 | 0.0633 / 0.1892 |
| 3 | 3 | 0.2579 / 0.5296 | 0.0800 / 0.2135 |
| 5 | 5 | 0.2135 / 0.4117 | 0.0913 / 0.2236 |
Table R4. Zero-shot accuracy on larger scenes.
| Method | Training M | M=10 ↓ | M=20 ↓ | M=50 ↓ |
|---|---|---|---|---|
| NEXUS | 2 | 0.1308 / 0.3464 | 0.1085 / 0.2756 | 0.0907 / 0.1395 |
| NEXUS | 3 | 0.1357 / 0.3709 | 0.1141 / 0.2747 | 0.1024 / 0.1331 |
| NEXUS | 5 | 0.1272 / 0.3454 | 0.1082 / 0.2677 | 0.0953 / 0.1283 |
| PNEO | 2 | 0.0838 / 0.2159 | 0.0762 / 0.2187 | 0.0825 / 0.1268 |
PNEO achieves lower RE@24 and RE@48 in every comparison. Against two-object-trained NEXUS, it reduces RE@48 by 37.7%, 20.7%, and 9.1% at M=10/20/50, respectively.
W2 (3). Computational cost and scaling
For the computational cost, please see our response to Q3 for runtime, peak inference memory, parameter counts, and object-count and point-cloud resolution scaling. We additionally report PNEO training costs:
| Training M | 2 | 3 | 5 |
|---|---|---|---|
| Full training time (h) | 16.7 | 18.6 | 26.4 |
| Peak training GPU memory (GB) | 0.68 | 1.31 | 3.29 |
Two-object training reduces training time by 36.7% and peak GPU memory by 79.3% relative to five-object training. The resulting model is reused at larger object counts without retraining.
Seed uncertainty. We additionally report PNEO results across three training seeds using the original training recipe, with each model trained only at M=2. Each cell below gives RE@24 / RE@48; the new results are reported as mean ± standard deviation.
| Test M | Paper number | Three-seed mean ± std |
|---|---|---|
| 2 | 0.0633 / 0.1892 | 0.0634 ± 0.0015 / 0.1824 ± 0.0092 |
| 3 | 0.0800 / 0.2135 | 0.0833 ± 0.0014 / 0.2041 ± 0.0043 |
| 5 | 0.0913 / 0.2236 | 0.0956 ± 0.0046 / 0.2226 ± 0.0050 |
| 10 | 0.0838 / 0.2159 | 0.0857 ± 0.0013 / 0.2198 ± 0.0052 |
| 20 | 0.0762 / 0.2187 | 0.0757 ± 0.0023 / 0.2178 ± 0.0111 |
| 50 | 0.0825 / 0.1268 | 0.0844 ± 0.0090 / 0.1286 ± 0.0017 |
Across M=2–50, all new RE@24 and RE@48 means differ from the submitted values by less than 5%, supporting the repeatability of PNEO's accuracy and object-count generalisation.
W2 (3) & Q3. Runtime, peak memory, parameter count and scaling
Thank you for requesting these measurements. Tables R5 and R6 report rollout execution time, peak inference memory, and parameter counts on NVIDIA GB10. Here we report time (s/trajectory) / memory (MB).
Table R5. Object-count scaling and baseline comparison.
| Method | Parameters (M) | M=2 | M=3 | M=5 | M=10 | M=20 | M=50 |
|---|---|---|---|---|---|---|---|
| PNEO (optimized; train M=2) | 0.676 | 0.120 / 138 | 0.143 / 139 | 0.218 / 141 | 0.532 / 148 | 1.772 / 168 | 1.872 / 293 |
| NEXUS (train M=2) | 0.352 | 2.114 / 89 | 2.324 / 97 | 3.057 / 114 | 6.010 / 157 | 17.919 / 525 | 98.545 / 2988 |
| GNS | 1.594 | 0.89 / 490 | 1.38 / 640 | 2.03 / 862 | 2.24 / 1,213 | 3.20 / 1,240 | 4.00 / 1,834 |
| MeshGraphNets | 3.853 | 1.52 / 95 | 1.79 / 119 | 2.34 / 190 | 3.66 / 261 | 6.61 / 468 | 15.73 / 1,035 |
| SGNN | 1.659 | 0.34 / 82 | 0.47 / 103 | 0.65 / 113 | 1.06 / 278 | 2.35 / 531 | 2.81 / 301 |
| EGNO | 0.219 | 0.24 / 444 | 0.44 / 878 | 0.75 / 1,702 | 1.69 / 4,207 | 4.09 / 15,538 | 2.35 / 2,500 |
| TIE | 0.769 | 0.33 / 49 | 0.47 / 65 | 0.99 / 114 | 3.48 / 345 | 12.26 / 1,258 | 68.73 / 7,615 |
| SBC | 0.623 | 0.45 / 116 | 0.60 / 170 | 0.86 / 241 | 1.82 / 438 | 3.91 / 787 | 23.81 / 1,085 |
| DPI-Net | 3.37–7.49 | 0.40 / 156 | 0.65 / 326 | 0.87 / 291 | — | — | — |
| EqCollide | 0.319 | 0.027 / 160 | 0.040 / 192 | 0.091 / 160 | — | — | — |
PNEO achieves the lowest reported execution time and peak inference memory among the measured baselines at M=20/50. At M=20/50, the reported rollout speedups over NEXUS are 10.1× to 52.6×, while the peak-memory ratio (NEXUS/PNEO) increases from 3.1× to 10.2×.
Table R6. Point-cloud resolution scaling at fixed M=2.
| Method | 64 | 128 | 256 | 512 | 1,024 |
|---|---|---|---|---|---|
| PNEO (optimized) | 0.100 / 105 | 0.118 / 106 | 0.167 / 108 | 0.313 / 112 | 0.656 / 119 |
| NEXUS (train M=2) | 2.038 / 81 | 2.099 / 89 | 2.293 / 107 | 2.378 / 147 | 2.868 / 207 |
| GNS | 0.50 / 226 | 0.65 / 339 | 1.50 / 728 | 3.87 / 1,447 | 9.42 / 2,718 |
| MeshGraphNets | 1.09 / 68 | 1.35 / 90 | 1.73 / 137 | 3.52 / 295 | 10.57 / 922 |
| SGNN | 0.15 / 40 | 0.29 / 66 | 0.49 / 120 | 1.22 / 254 | 2.88 / 469 |
| EGNO | 0.06 / 94 | 0.19 / 265 | 0.67 / 1,031 | 2.52 / 4,387 | 10.80 / 20,329 |
| TIE | 0.16 / 39 | 0.28 / 49 | 0.62 / 87 | 2.36 / 235 | 8.47 / 820 |
| SBC | 0.33 / 60 | 0.43 / 91 | 0.70 / 186 | 1.44 / 441 | 3.29 / 839 |
| DPI-Net | 0.27 / 64 | 0.34 / 108 | 0.71 / 329 | 2.07 / 1,329 | 3.83 / 2,556 |
PNEO is faster than NEXUS across all tested resolutions. Its memory advantage is most pronounced at 1,024 points per object, where it achieves a 4.4× rollout speedup and uses 119 MB peak memory, compared with 207 MB for NEXUS.
Thank you for your response. I will raise my score to 3.
W1 & Q1. Energy interpretation, velocity conditioning, energy balance and momentum
We agree that the derived force does not strictly correspond to the full gradient of the state-dependent potential, whose coefficients depend on both positions and velocities. However, we adopt this formulation as an engineering choice rather than a strict implementation of conservative dynamics, and provide additional ablation results to assess its empirical benefits. We will revise the manuscript to clarify this distinction and qualify the corresponding conservative-energy claims.
Including velocity information allows the encoder to capture richer collision-state information. To assess its contribution, we compare PNEO with a position-only variant that excludes velocity from the potential branch while retaining it in the dissipation branch. Both variants use the same architecture and coefficient-freezing convention and are trained at (M=2). We evaluate them at (M=2,3,5), with each table entry reporting RE@24 / RE@48. Velocity-conditioned PNEO achieves lower errors at both horizons across all three object counts in this comparison, providing empirical support for this architectural choice.
To quantify the energy changes caused by coefficient updates, we explicitly measure the energy balance:
ET−E0=Wext−D+Jc+Rnum,
where E is total mechanical energy (kinetic, gravitational, structural and contact energies), E0=E(0), D is cumulative damping dissipation, and Wext is external work. Jc accumulates stored-energy changes caused by coefficient updates at fixed positions and velocities, where positive values indicate energy injection. Rnum denotes the numerical energy-balance residual.
Table R1. Energy balance. Values are trajectory means normalized by E0. The physical reference uses the fixed-coefficient simulator.
Energy balance. As shown in Table R1, coefficient updates add energy equivalent to 1.2% of E0 on average by the final state, while the numerical balance residual is −0.9% of E0. This explicitly quantifies the energy change caused by coefficient updates without assuming exact conservation.
Table R2. Momentum and collision impulses. Values are trajectory means normalized by Pref=2mtotE0, where mtot is the total dynamic mass. This normalization makes the measurements dimensionless and enables comparison across trajectories with different total masses and initial energies. Negative vertical momentum indicates downward motion.
Momentum balance. We additionally check whether momentum changes match the applied external impulses. Table R2 reports trajectory means normalized for total mass and initial energy, showing near-zero net inter-object impulses and balance residuals. The maximum normalized residual across all trajectories and frames is 1.4×10−7.