{"id":"1249c2e6-8bd4-47ed-9a24-63929659d1cf","arxiv_id":"2411.11467","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"HOPNet uses combinatorial complex topology plus a physics-inspired message schedule to learn rigid body collisions, reporting lower rollout error and better OOD generalization than graph-based baselines.","lead":"HOPNet is a neural network that learns rigid body collisions by representing objects, surfaces, and contacts as higher-order topological cells and passing messages along physics-inspired pathways. It reports lower position and orientation errors over long rollouts than graph-based baselines, plus generalization to unseen shapes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on an uncontrolled FIGNet comparison: original published numbers plus a supplement-only reimplementation, with possible feature mismatch (Eq. 6); reproduction is needed before the margin is trusted.","rationale":"The reader's weakest assumption—unfair or unfaithful FIGNet comparison—is the same load-bearing concern I identify. I add one concrete dimension: HOPNet's Eq. 6 gives the model explicit mass, friction, and restitution coefficients, so feature parity with FIGNet_reimpl must be verified before the margin can be attributed to the architecture. The paper provides code and data links, which is good support, but no formal verification or executed reproduction was reviewed. The reported results are credible yet conditional, so the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":18466,"tokens_out":6651,"duration_ms":70504,"concrete_test":"Run the released HOPNet code with the provided FIGNet_reimpl configuration, replacing the reported FIGNet numbers by a single controlled baseline: train FIGNet_reimpl on the exact MOVi-A/B splits used by HOPNet with the same physical-parameter inputs (including mass, friction, restitution if used) and the same number of training steps, then recompute Fig. 4. If the controlled baseline matches the original FIGNet numbers within seed noise and HOPNet still beats it by the reported margin at T=50 and T=75, the concern is settled; if not, the SOTA claim must be re-benchmarked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that HOPNet is substantially more accurate than FIGNet over 50-step rollouts and that this advantage extends to 75 steps—rests on a comparison that is not fully controlled in the paper. Figure 4 mixes two sources: (i) the original FIGNet/MeshGraphNet numbers taken from [6], which may have been produced on different MOVi data splits, environment configurations, or evaluation protocols; and (ii) a FIGNet_reimpl that is described only in Supplement 2.5. Because the original FIGNet numbers are not reproduced under HOPNet's exact conditions, the 'substantial margin' could be partly a dataset/protocol artifact rather than an architectural improvement. A second, more specific confound is feature asymmetry: HOPNet's object features h(4) (Eq. 6) explicitly include mass, friction coefficient c1, and restitution coefficient c2. The paper does not state whether FIGNet_reimpl was given the same privileged physical parameters. If the baseline lacked these inputs (or received them in a different form), the comparison would not isolate the proposed topology/physics-informed message passing. This does not make the architecture's internal consistency doubtful, but it is the load-bearing point for the SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces HOPNet, a topological message-passing network for learning rigid-body dynamics from mesh representations. It represents an environment as a combinatorial complex whose cells correspond to nodes, edges, mesh triangles, collision contacts, and objects, and it orders message passing in a sequence motivated by Newtonian mechanics (collisions first, then object-level updates). Experiments on the Kubric MOVi-spheres/A/B/C datasets report lower position and orientation RMSE after 50 rollout steps than MeshGraphNet and FIGNet, improved long-horizon stability, OOD generalization to unseen mesh geometries, counterfactual rollouts, and ablation studies of the proposed components.","tokens_in":18733,"tokens_out":10722,"duration_ms":110761,"significance":"If the reported gains are confirmed under a fully controlled protocol, HOPNet would be a useful advance in learned rigid-body simulation on mesh inputs: it is one of the first demonstrations that higher-order/combinatorial complex representations combined with collision-ordered message passing can beat graph-network baselines on multi-object benchmarks. Strengths of the paper include public code/data, error bars over three seeds for HOPNet, explicit OOD evaluation, and ablation studies isolating the topological cells, the center-of-mass feature, and the sequential message-passing order. The main open issue is the fairness and completeness of the FIGNet comparison.","major_comments":[{"comment":"The central comparison against MeshGraphNet and FIGNet is not fully controlled. The figure combines numbers taken directly from [6] with results from a reimplementation (FIGNet_reimpl) whose details are only in Supplementary Section 2.5; the published numbers may have been obtained under different data splits, object scaling, or evaluation protocols, and no uncertainty is reported for them. Because the abstract's claim of 'superior accuracy' is based on this comparison, the authors should either reproduce all baselines under the exact conditions used for HOPNet or explicitly restrict the state-of-the-art claim to the controlled reimplementation comparison.","section":"Section 3.1, Figure 4"},{"comment":"The paper does not state whether the FIGNet_reimpl baseline received the same privileged physical parameters that HOPNet uses: mass m, friction coefficient c1, and restitution coefficient c2 in the object features h^(4) (Eq. 6). If FIGNet_reimpl did not have access to these parameters, the comparison is not architecture-vs-architecture. Please report the full feature set of the reimplementation and, if it differs, run a variant with matched inputs.","section":"Section 3.1, Eq. (6)"},{"comment":"The quantitative claim that HOPNet performs '50% more rollout steps' before reaching the same error as the best baseline is not supported by curves in the main text and again relies on the uncontrolled baseline numbers from [6]. The full error-versus-horizon curves for all methods under identical conditions should be shown, or the claim should be removed.","section":"Section 3.1, paragraph beginning 'Importantly'"}],"minor_comments":[{"comment":"The text refers to 'Figure 8a', 'Figure 8b', and 'Figure 8c' when discussing the ablations; these should be 'Figure 9a', 'Figure 9b', and 'Figure 9c'.","section":"Section 3.4, Figures 8 and 9"},{"comment":"The symbol m3->2_s->r is used in Eq. (11), but the text below the equation calls it m2->3_s->r; please make the notation consistent.","section":"Eqs. (10)-(11) and surrounding text"},{"comment":"The phrase 'To ensure energy conservation' overstates what is implemented: no conservation constraint is enforced in the learned messages. Suggest rewording to 'motivated by' or 'designed to encourage'.","section":"Section 5.2.2"},{"comment":"The phrase 'This counterintuitive results' should be 'This counterintuitive result'.","section":"Section 3.2"},{"comment":"The phrase 'trained on exactly the same datasets than HOPNet' should be 'the same datasets as HOPNet'.","section":"Section 3.1"},{"comment":"It would be clearer to state explicitly which bars come from [6] and which are from the reimplementation, since the legend includes both FIGNet and FIGNetreimpl.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the main risk is the uncontrolled baseline comparison; the reimplementation is described only in the supplement. I would encourage requesting the supplement details and possibly a matched-feature baseline before publication. The architecture and internal consistency of the paper appear sound, and the authors' provision of code and data is a positive signal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHOPNet is a real contribution: the first (as far as I know) use of combinatorial complex message passing for rigid-body dynamics, with distinct cell types for triangles, collision contacts, and objects, and a sequential message schedule that mirrors collision-then-propagation. The ablations are clean—removing object-level cells, center-of-mass distance, and sequential scheduling all degrade performance as expected—and the authors ship code and data, which makes the work reproducible in principle. The OOD experiments, especially training on spheres and transferring to MOVi-C, are genuinely informative.\n\nThe soft spot is one the authors themselves expose: the SOTA comparison to FIGNet rests partly on numbers from the original paper and partly on a reimplementation described only in the supplement. That is a common situation when no official code exists, and the reimplementation baseline weakens the main concern, but it does not eliminate it. I would want to see the supplement's Section 2.5 and confirm that FIGNet_reimpl was given the same object-level features (mass, friction, restitution) as HOPNet before trusting the margin in Figure 4. If that check passes, the central claim likely holds. A smaller phrasing issue: \"ensures energy conservation\" in Methods 5.2.2 overstates what permutation-invariant learned messages do; they only bias the model toward symmetric updates.\n\nNone of this is fatal. The architecture is sensible, the evaluation is thorough, and the limitations (scaling to millions of faces, reliance on explicit state) are stated plainly. This is a paper for people working on learned physics simulation or topological deep learning; it will be useful to them as a baseline and a design source. I would send it to a serious referee: the review should be conditional on the FIGNet comparison being pinned down, but the work itself deserves the time.","headline":"A genuinely new architecture for learned rigid-body dynamics with clean ablations; the SOTA claim is plausible but hinges on a FIGNet comparison that needs supplement verification.","tokens_in":19214,"tokens_out":2079,"would_cite":false,"duration_ms":21112,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A topological network with physics-guided messages learns rigid-body collisions more accurately than graph simulators.","keywords":["rigid body dynamics","topological deep learning","combinatorial complex","physics-informed message passing","graph neural networks","collision simulation","out-of-distribution generalization","autoregressive rollout"],"falsifier":"Retrain HOPNet and a faithful FIGNet reimplementation from the same codebase, data splits, random seeds, and compute budget, then compare 50-step RMSE on MOVi-A; if the margin disappears or reverses, the central claim fails. A second check: if HOPNet's translation RMSE at step 75 exceeds FIGNet's published 50-step value on MOVi-A, the claimed rollout-length advantage is falsified.","tokens_in":18239,"feed_emoji":"⚙️","tokens_out":7994,"duration_ms":68604,"temperature":0.7,"pith_summary":"Graph-based simulators represent a rigid-body scene as nodes and edges, which discards the surfaces, the objects, and the contact geometry that actually determine collisions. The paper proposes HOPNet, which instead represents each scene as a combinatorial complex with five ranks of cells—nodes, edges, mesh triangles, collision contacts, and objects—and exchanges learnable messages along pathways chosen from Newtonian mechanics. On the MOVi-A and MOVi-B benchmarks, HOPNet reports lower position and orientation error than MeshGraphNet, FIGNet, and a same-split FIGNet reimplementation after 50 autoregressive rollout steps, and it stays below the best baseline's 50-step error for about 75 steps. The authors further show that a model trained only on spheres generalizes to complex and high-resolution unseen meshes, which they interpret as evidence that the model learns transferable face-to-face collision dynamics rather than object-specific patterns. If the claims hold, learned physics simulation should move from node-level graphs to higher-order topological representations with physically structured message passing.","feed_headline":"Physics-guided topology beats graph nets at rigid-body simulation","feed_subtitle":"HOPNet reports lower rollout error on MOVi-A and MOVi-B and keeps errors low on unseen high-resolution meshes.","key_machinery":"The load-bearing object is the spatiotemporal combinatorial complex $\\mathcal{X}^t$, a hierarchy of cells of five ranks—nodes $\\mathcal{X}^{(0)}$, edges $\\mathcal{X}^{(1)}$, mesh triangles $\\mathcal{X}^{(2)}$, collision contacts $\\mathcal{X}^{(3)}$, and objects $\\mathcal{X}^{(4)}$—each with its own learned features. The argument is carried by a sequential physics-informed message-passing scheme in the Encode-Process-Decode style: faces are first enriched with intra-object node, edge, and object messages; collisions are then processed at contact cells built from the closest points between triangles of different objects; and the collision effects are propagated back through objects to nodes before accelerations are decoded. The final pose comes from a second-order forward-Euler integration of node accelerations followed by shape matching, which enforces rigidity. This machinery is what lets the model treat collisions as surface-to-surface momentum exchanges rather than node-to-node events, and the ablations attribute the accuracy gains to exactly these components.","core_discovery":"The central claim is that the information lost by graph representations—triangle-level surface structure, object-level cohesion, and explicit collision contacts—is exactly what a neural simulator needs to learn rigid-body collisions, and that a topological message-passing network can exploit it. HOPNet encodes the whole scene as a spatiotemporal combinatorial complex $\\mathcal{X}^t$ with cells of ranks 0 through 4: nodes, edges, mesh triangles, collision contacts, and objects. Its message-passing proceeds in a fixed Newtonian order: enrich triangle embeddings with node, edge, and object information; compute collision effects between contacting triangles using the closest-point geometry of each face; propagate the collision updates back to objects and then to nodes; and decode per-node and per-object accelerations, which are integrated with second-order forward Euler and shape matching to produce the next pose. The paper reports that this design outperforms both graph baselines on MOVi-A and MOVi-B, that removing the object-level cells raises translation error by 44.4 percent, that removing the node-to-center-of-mass distance raises error by 26 percent on MOVi-A, and that replacing the sequential physics-guided message flow with unrestricted message passing raises error by 14 percent even with 34 percent more parameters.","pith_inferences":["The same rank-structured message flow could be applied to deformable bodies or fluids by replacing rigid shape matching with a softer positional update, an extension the paper itself lists as future work.","Since the collision message is computed from local triangle geometry rather than object identity, a direct test with unseen object counts and mixed mesh resolutions would probe how far the universal collision dynamics claim extends.","Coupling HOPNet with a perception module that estimates mass, friction, and restitution from images would remove the current dependence on explicit state inputs and make the approach usable on raw video.","The small loss from cutting 74 percent of the parameters suggests representational structure, not raw capacity, drives performance, so hierarchical or adaptive mesh message passing may extend accuracy before model size becomes the bottleneck."],"forward_implications":["On MOVi-A and MOVi-B after 50 rollout steps, HOPNet reports lower position and orientation RMSE than MeshGraphNet, FIGNet, and a same-split FIGNet reimplementation.","The model performs about 50 percent more rollout steps than the best baseline before reaching the same translation error, reaching FIGNet's 50-step RMSE only at step 75.","A model trained only on spheres generalizes to MOVi-A, MOVi-B, and the high-resolution scanned meshes of MOVi-C, and masking slow collisions during training improves generalization further.","Ablations show that object-level cells, the node-to-center-of-mass distance feature, and the sequential physics-guided message flow each contribute substantially to accuracy.","Because the representation is rebuilt at every timestep, the model can accommodate counterfactual interventions such as removing an object or changing initial conditions, and performance stays stable when the collision radius is halved or doubled."],"supporting_citations":[{"why":"Provides the MeshGraphNet baseline whose reported rollout errors HOPNet compares against on MOVi-A and MOVi-B.","marker":"[5]"},{"why":"Provides the FIGNet baseline, the published comparison numbers, and the reference implementation that the same-split reimplementation extends to the MOVi datasets.","marker":"[6]"},{"why":"Supplies the Encode-Process-Decode paradigm and the particle-based graph-network simulation approach that HOPNet builds on.","marker":"[23]"},{"why":"Establishes the message-passing graph-network framework and relational inductive biases that the paper contrasts with its physics-informed topological message passing.","marker":"[27]"},{"why":"Defines the combinatorial-complex and topological message-passing formalism HOPNet adopts for cells of different ranks.","marker":"[37]"},{"why":"Supplies the shape-matching procedure used to recover rigid object poses from predicted node positions.","marker":"[42]"},{"why":"Provides the benchmark datasets MOVi-spheres, MOVi-A, MOVi-B, and MOVi-C used in all experiments.","marker":"[43]"},{"why":"Provides the high-resolution scanned household-object meshes used in MOVi-C for the extreme out-of-distribution generalization test.","marker":"[44]"}],"fun_headline_variants":["Physics+topology network beats graph nets on rigid-body sims","HOPNet: topology and physics cut rigid-body rollout errors","Why graph nets fail at rigid-body sims: topology fixes it","Higher-order topology helps neural nets learn collisions better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against FIGNet is the load-bearing premise: HOPNet's reported margin relies on the FIGNet numbers published in the original paper and on a reimplementation trained on the same splits, and if those baselines were produced under different conditions the advantage could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Physics+topology network beats graph nets on rigid-body sims","HOPNet: topology and physics cut rigid-body rollout errors","Why graph nets fail at rigid-body sims: topology fixes it","Higher-order topology helps neural nets learn collisions better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1328,"prompt_tokens":952,"completion_tokens":376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":568,"tokens_out":376,"duration_ms":16384,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:29:39.662969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain HOPNet and a faithful FIGNet reimplementation from the same codebase, data splits, random seeds, and compute budget, then compare 50-step RMSE on MOVi-A; if the margin disappears or reverses, the central claim fails. A second check: if HOPNet's translation RMSE at step 75 exceeds FIGNet's published 50-step value on MOVi-A, the claimed rollout-length advantage is falsified.","supporting_citations":[{"cited_title":"& Gross, M","cited_arxiv_id":null,"evidence_quote":"Supplies the shape-matching procedure used to recover rigid object poses from predicted node positions."},{"cited_title":"& Battaglia, P","cited_arxiv_id":null,"evidence_quote":"Provides the MeshGraphNet baseline whose reported rollout errors HOPNet compares against on MOVi-A and MOVi-B."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FIGNet baseline, the published comparison numbers, and the reference implementation that the same-split reimplementation extends to the MOVi datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Encode-Process-Decode paradigm and the particle-based graph-network simulation approach that HOPNet builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark datasets MOVi-spheres, MOVi-A, MOVi-B, and MOVi-C used in all experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the high-resolution scanned household-object meshes used in MOVi-C for the extreme out-of-distribution generalization test."}],"review_version":1}