{"id":"6b78e546-2a6e-4e96-bb05-18b42e09a8ba","arxiv_id":"2505.01059","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Model Tensor Planning uses random multipartite graph sampling with spline interpolation to increase exploration in sampling-based MPC, outperforming several baselines in simulated robot tasks.","lead":"This paper proposes Model Tensor Planning, a sampling-based model predictive control method that generates diverse control candidates by sampling paths through random graphs and smoothing them with splines. It is a candidate improvement for robots that need to explore many possible actions quickly, such as dexterous hands and humanoid robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theory-experiment gap: Theorem 1 is linear-interpolation and asymptotic only, while Table 3 evaluates M=2–3 spline variants; at M=2 the sampler support collapses to a low-dimensional affine manifold, so the asymptotic coverage and maximum-entropy claims do not support the reported exploration…","rationale":"The reader's weakest assumption and my independent reading converge on the same load-bearing concern: the paper's only theoretical guarantee (Theorem 1) is asymptotic and linear-interpolation-specific, while the deployed algorithm uses small-M spline variants, and no finite-M or spline-aware bound is provided. I sharpen this in two ways. First, the gap is structural rather than just quantitative: at M=2, B-spline and Akima interpolation both collapse to straight lines between two waypoints, so the sampler's support is a finite-dimensional, measure-zero submanifold of the trajectory space. Second, the 'maximum entropy' statement in the abstract and Appendix A.2 mixes discrete and differential entropy, so it is not a well-posed mathematical claim. These issues do not by themselves invalidate the experimental results: MTP may still work well because low-dimensional smooth splines are a good prior for the tested tasks, and the paper's ablations give some empirical support for this. But the central conceptual claim that MTP achieves global exploration via asymptotic path coverage and maximum entropy is unsupported as stated. Since the reader already assigned CONDITIONAL with moderate confidence and this was identified as the weakest assumption, my stress-test does not move the verdict; it strengthens the reason for treating the theory as a guiding heuristic rather than a guarantee. The concrete test would empirically check whether the finite-M manifold used in the experiments actually covers task-relevant control trajectories, thereby determining whether the concern is primarily a presentation issue or a substantive limitation of the method's exploration mechanism.","tokens_in":21861,"tokens_out":11069,"duration_ms":128154,"concrete_test":"Compute, for each task where MTP succeeds (e.g., G1-Standup, G1-Walk, PushT), the normalized sup-norm distance from reference or expert control trajectories to the M=2 B-spline/Akima manifold: project each reference trajectory onto the best two-waypoint spline (closed-form for linear interpolation, least squares for B-spline with the paper's fixed knots) and report the residual relative to control limits. If the residual is large, the evaluated MTP configurations cannot generate the demonstrated good controls through tensor sampling alone, so the reported advantage is not explained by Theorem 1; if the residual is small, the low-dimensional manifold happens to cover the task and the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central support for MTP's exploration claim rests on Theorem 1 (Section 3.2) and the entropy analysis (Appendix A.2), both stated only for linear interpolation in the limit M,N→∞. Section 3.1 explicitly defers B-spline and Akima analyses to future work, yet all main experiments use B-spline (p=2) or Akima interpolation with M=2–3 (Table 3). This gap is not merely quantitative: for M=2, both interpolation schemes reduce to the line segment between the two random waypoints (a two-control-point B-spline is an affine combination, and a two-point Akima spline has zero quadratic and cubic coefficients). The support of tensor sampling is then a 2n-dimensional affine submanifold of the T×n control-trajectory space, which has measure zero for any finite M. For G1-Walk (n=29, T=40), that is a 58-dimensional set inside a 1160-dimensional space. Consequently the 'asymptotic path coverage' theorem provides no finite-M or spline-specific denseness guarantee, and the paper's own M,N sweep (Fig. 8) shows performance degrades as M grows, consistent with a low-dimensional prior rather than dense coverage. The 'maximum entropy' claim is also internally shaky: Appendix A.2 compares a discrete entropy M log N over N^M index sequences with the differential entropy of a Gaussian (Eq. 22), which are not commensurable quantities. The empirical wins may therefore reflect the favorable inductive bias of low-degree splines, not a validated high-entropy exploration mechanism over control-trajectory space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Model Tensor Planning (MTP), a sampling-based MPC framework in which control candidates are generated by uniformly sampling waypoints on each layer of a random multipartite graph, interpolating the resulting point sequences (linearly, with B-splines, or with Akima splines), and mixing a fraction beta of these global samples with local Gaussian samples inside a CEM/MPPI-style update. The theoretical section claims asymptotic path coverage and maximum entropy of the tensor sampler as the number of graph layers M and waypoints per layer N tend to infinity. Experiments on six sim-to-sim MuJoCo tasks compare MTP variants against MPPI, Predictive Sampling, OpenAI-ES, and Diffusion Evolution, and additional ablations study beta, elite count, M, N, and B-spline degree, together with JAX/MuJoCo XLA wall-clock benchmarks.","tokens_in":22232,"tokens_out":11469,"duration_ms":127166,"significance":"If the theoretical claims were properly scoped, MTP would be a useful contribution to sampling-based MPC: the beta-mixing idea is simple and practical, the JAX implementation with JIT/vmap and batched rollouts is a genuine engineering strength, and the paper provides planning-time benchmarks and careful ablations of beta and E. The broad empirical comparison against both MPC and evolutionary baselines is also a strength. However, the theoretical support for the central 'high-entropy exploration' claim currently applies only to linear interpolation in an asymptotic regime, while the evaluated algorithm uses spline variants at small M; the maximum-entropy argument is also not formally valid. The empirical results may still stand on their own, but the paper's central theoretical narrative needs substantial revision.","major_comments":[{"comment":"The proof of Theorem 1 is internally inconsistent and incomplete. It states 'Since U is open' despite the paper's earlier assumption that the control space is compact; if u(t_i) lies on the boundary of U, an ambient ball is not contained in U (relative balls would fix this). More importantly, the theorem is stated as a deterministic limit, but the proof establishes only a probabilistic statement: the union bound gives success probability at least 1 - M e^{-cN}. For fixed M this tends to 1 as N grows, but under the double limit M,N -> infinity one must impose a rate such as N = omega(log M), which is absent. As written, the deterministic claim 'lim_{M,N->infty} min_g ||u-g||_infty = 0' is not established.","section":"Appendix A.1, Theorem 1"},{"comment":"The path-coverage guarantee is proved only for linear interpolation, and Section 3.1 explicitly defers B-spline and Akima analyses to future work, yet Table 3 reports M=2-3 for the spline variants used in all main comparisons. For M=2, a degree-2 B-spline over two control points is an affine segment, and the Akima construction in Eq. (9) references m_{i-2}, so the M=2 Akima case is not even covered by Definition 3 as written. The support of tensor sampling is then a 2n-dimensional affine manifold inside the T x n trajectory space (e.g., dimension 58 versus 1160 for G1-Walk), not a dense subset. Consequently Theorem 1 does not support the exploration claim for the algorithm actually evaluated; the paper needs a finite-M or spline-specific denseness bound, or a clear restriction of the theoretical claims to the linear-interpolation asymptotic setting.","section":"Section 3.1-3.2, Table 3"},{"comment":"The maximum-entropy argument compares incommensurable quantities. H(P_G) = M log N is the Shannon entropy of a finite categorical distribution over N^M index sequences, while H(P_MPPI) in Eq. (22) is a differential entropy in nats for a Gaussian over continuous control trajectories. The divergence of M log N as M,N grow reflects an increasingly fine discretization, not convergence to a maximum-entropy distribution over the continuous trajectory space, and there is no reference measure on the infinite-dimensional path space that would make the abstract's 'maximum entropy in the control trajectory space' claim meaningful. The maximum-entropy statement is therefore unsupported as written.","section":"Appendix A.2, Eqs. (20)-(22)"}],"minor_comments":[{"comment":"For beta = 1, the expression B - P - 1 is negative, yet beta = 1 is used in the Navigation experiments and in the Fig. 5 sweep; the paper should specify max(B - P - 1, 0) or sample B - P - 1 local trajectories only when beta < 1.","section":"Algorithm 2, Line 7"},{"comment":"The control waypoint tensor is written as Q in R^{M x N x d}, but the control dimension is n throughout the paper; this should be R^{M x N x n}.","section":"Algorithm 2, Line 4"},{"comment":"The standard CEM covariance update divides by E - 1, which is undefined for E = 1; the ablation in Fig. 5 includes E = 1, so the preliminary equation should either state E > 1 or explain how the softmax/min-variance update in Algorithm 2 avoids this case.","section":"Section 2.1, Eq. (4)"},{"comment":"If Sigma is the per-step covariance, the differential entropy of the Gaussian path distribution should contain (T/2) log det(Sigma), not T log det(Sigma); the factor of two affects the comparison and should be corrected.","section":"Appendix A.2, Eq. (22)"},{"comment":"Calling the graph a complete M-partite directed graph is imprecise: edges only connect consecutive layers, not arbitrary pairs across partitions; 'consecutive-layer complete directed bipartite graph' would be more accurate.","section":"Definition 1"},{"comment":"The text says sampling without replacement is O(NM), but the number of combinatorial paths is N^M; the intended complexity is presumably O(N^M), and the notation should be fixed.","section":"Section 3.1"},{"comment":"The statement that for M,N -> infinity any g in F is also in G(M,N) is not well-defined, because the limiting object G(infinity, infinity) is not defined; the remark should be rephrased in terms of the convergence of the minimum distance.","section":"Section 3.2, Remark"},{"comment":"The y-axis label 'PustT' is a typo for 'PushT'.","section":"Appendix A.5, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The overlap with the prior GTMP paper (Le et al. 2025) is clearly cited, and the MPC-specific contribution, especially the beta-mixing update and the spline-based tensor sampler, is distinct enough. My main concern is that the theoretical claims as stated are broader than what the proofs actually support; this is fixable by scoping the theorems to linear interpolation, making the probabilistic statement and rate explicit, and correcting the entropy discussion, but it requires more than cosmetic changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MTP is a genuinely useful engineering contribution to sampling-based MPC, but its headline theoretical guarantee doesn't apply to the spline variants that actually win the experiments. Read it for the algorithm and ablations, not for the asymptotic theory.\n\nWhat's new: the idea of using random multipartite graph sampling (from the authors' GTMP) as a high-entropy proposal distribution inside a CEM/MPPI loop, with beta-mixing between global tensor samples and local Gaussian samples, and B-spline/Akima interpolation. The implementation is clean, fully vectorized in JAX/MuJoCo XLA, and the paper reports real-time planning benchmarks. The ablations are thorough—beta sweep, elite count, M/N sweep, spline degree—and show the design choices matter in predictable ways. Empirically, MTP variants consistently beat MPPI, PS, OpenAI-ES, and Diffusion Evolution on contact-rich tasks like PushT and Cube-In-Hand, and match them on simple tasks like Walker. That's a real result.\n\nSoft spots:\n\n1. The theory-experiment gap is significant. Theorem 1 is proved only for linear interpolation and in the limit M,N → ∞. The paper explicitly defers B-spline and Akima analysis to future work. Yet every main experiment uses B-spline (p=2) or Akima with M=2 or 3. At M=2, both interpolation schemes give the straight line between two waypoints, so the support is a low-dimensional affine subspace—measure zero in control trajectory space. The claimed \"asymptotic path coverage\" and \"maximum entropy\" therefore do not actually explain the exploration behavior that makes MTP work in practice. The empirical success is probably due to the favorable inductive bias of low-degree splines and the beta-mixing schedule, not a validated high-entropy coverage mechanism.\n\n2. The entropy discussion in Appendix A.2 compares discrete entropy (M log N) of graph paths with differential entropy of a Gaussian. Those are different quantities; the comparison doesn't establish maximum entropy over control trajectories.\n\n3. The baseline comparison is a bit tilted: MTP hyperparameters are tuned per task while the evolutionary baselines use default evosax parameters. That's common in this literature, but it means the \"outperforms ES\" claim is softer than it looks.\n\n4. No code/data released, though the setup is sim-to-sim and hydrax-based, so reproducible in principle.\n\nThese are fixable. The paper would be stronger if it either (a) added a finite-M concentration or explicit spline-manifold result, or (b) dropped the asymptotic framing and presented tensor sampling as a structured low-dimensional prior with empirical exploration benefits.\n\nWho it's for: anyone working on sampling-based MPC for contact-rich or high-dimensional tasks, especially with GPU simulators. It deserves a serious referee—the algorithm is practical, the experiments are extensive, and the gap between theory and practice is worth an editor's attention rather than a desk reject.\n\nRecommendation: send it to review, but the reviewers should push for either a corrected theory section or a more honest framing.","headline":"MTP is a genuinely practical extension of the authors' GTMP tensor-sampling idea into sampling-based MPC, with strong empirical results on contact-rich tasks; the catch is that the paper's theoretical guarantee, which is only for linear interpolation in an asymptotic limit, does not cover the spline variants that actually produce the wins.","tokens_in":22777,"tokens_out":2650,"would_cite":true,"duration_ms":24290,"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":"Tensor sampling frees model-predictive control from local minima","keywords":["model predictive control","tensor sampling","multipartite graphs","cross-entropy method","B-spline interpolation","Akima spline","exploration exploitation trade-off","GPU-accelerated planning"],"falsifier":"Construct a control task with a narrow cost basin that can only be reached by a specific smooth trajectory whose waypoint sequence is not captured by the small $M,N$ grid. If MTP with $M=2-3$, $N=50-100$ and B-spline or Akima interpolation cannot concentrate samples there while a linear-interpolation variant or a denser graph can, the coverage claim for practical settings fails. More directly, compute the distance from a family of spline-interpolated paths to the nearest linear-interpolated graph path for the $M,N$ values used in Table 3; if the distance does not shrink toward zero or if the spline paths fail to intersect the low-cost region, the transfer assumption is false.","tokens_in":21668,"feed_emoji":"🤖","tokens_out":3613,"duration_ms":31950,"temperature":0.7,"pith_summary":"Sampling-based model predictive control is powerful for nonlinear and contact-rich robots, but standard methods sample noise around a nominal trajectory, which traps them in local minima. This paper proposes Model Tensor Planning (MTP), which reformulates control sampling as tensor operations over randomized multipartite graphs: it uniformly samples control waypoints organized into layers, then interpolates smooth trajectories through them with B-splines or Akima splines. The paper claims this structured sampling achieves high-entropy, globally diverse control candidates at the same GPU cost as standard samplers, and that a beta-mixing blend of these global samples with local refinement in the Cross-Entropy update balances exploration and exploitation. If correct, it would make sampling MPC far more exploratory without dense reward shaping or differentiable dynamics, while remaining real-time and vectorized.","feed_headline":"Tensor sampling frees MPC from local minima","feed_subtitle":"Control paths sampled over random graphs match or beat evolutionary baselines in dexterous manipulation and humanoid locomotion.","key_machinery":"The central machinery is tensor sampling over a random multipartite graph $G(M,N)$ control discretization: $M$ layers each containing $N$ control waypoints drawn uniformly from the control box, with complete directed connections between consecutive layers. Sampling with replacement from each layer forms a batch tensor $C \\in \\mathbb{R}^{B \\times M \\times n}$, which is interpolated along the time axis into full-horizon control trajectories using either linear segments, degree-$p$ B-splines (whose basis functions form a convex combination of waypoints), or Akima splines (piecewise cubics through the waypoints with local slope averaging). The graph structure converts path diversity into cheap tensor index operations, and the asymptotic path-coverage theorem converts the denseness of the graph in the limit into a promise that the sampler can reach any feasible control path.","core_discovery":"The central claim is that a control trajectory can be treated as a path through a random multipartite graph whose layers are uniform random samples of control space, with each layer holding N candidate waypoints and edges connecting consecutive layers; sampling paths by picking one waypoint per layer and interpolating yields globally diverse, smooth control trajectories. The paper proves (Theorem 1) that for any bounded-variation control path, the distance from that path to the nearest graph path goes to zero as the number of layers M and waypoints per layer N both go to infinity, so the graph asymptotically covers the space of feasible control paths. Since each layer is sampled uniformly and independently, the distribution over graph paths has entropy $M \\log N$, which grows without bound, meaning the sampler approaches maximum entropy over trajectory space. The method embeds these tensor-sampled trajectories into a modified Cross-Entropy update that softmax-weights elite candidates and mixes global tensor samples with local Gaussian samples in proportion $\\beta$; experiments on dexterous manipulation and humanoid locomotion claim MTP matches or outperforms MPPI, Predictive Sampling, OpenAI-ES, and Diffusion Evolution in task success and control robustness, with ablations showing moderate $\\beta$ values and B-spline/Akima interpolation are important.","pith_inferences":["The asymptotic path-coverage guarantee is proven only for linear interpolation and in the infinite limit; the paper's own experiments rely on B-spline and Akima interpolation at $M=2$ to $3$, so an implicit assumption is that finite-layer spline interpolation of graph paths still covers the cost-relevant control regions. A finite-$M$ or spline-specific denseness bound would directly test this.","The entropy identity $H = M \\log N$ suggests a concrete scaling rule for sample efficiency: if the rollout budget $B$ is fixed, increasing $M$ multiplies the path space exponentially, so the effective coverage per sampled path shrinks; the paper's own recommendation to scale $N$ with $B$ is a partial answer, and a coverage-versus-budget analysis could make that tradeoff precise.","Because the tensor sampler is independent of the cost function and dynamics, it could be plugged into other trajectory optimizers or used as a high-entropy proposal distribution for diffusion-based or learned planners without changing the planner itself.","The $\\beta$-mixing strategy could be made adaptive per timestep, using the entropy or cost variance of the batch to raise $\\beta$ when the planner is stagnant and lower it when following an established nominal trajectory, which is a direct extension the paper leaves implicit."],"forward_implications":["MPC on GPU accelerators could maintain both global exploration and local refinement within a single batched rollout, replacing the usual choice between exploratory noise and exploitation.","Trajectory entropy becomes a tunable design quantity: increasing $M$ or $N$ raises the entropy of the candidate set, with the $M \\log N$ closed form guiding how to spend a fixed rollout budget.","Because the sampler does not rely on gradients or dense reward shaping, it may extend to partially observed or contact-rich platforms where differentiable dynamics and informative costs are unavailable.","B-spline and Akima interpolation decouple the coarse waypoint exploration from the smoothness of the executed control, so exploration can be aggressive while the executed trajectory remains feasible.","The same tensorized graph structure, with vectorized JIT rollouts, suggests that online domain randomization over many dynamics models can be folded into the sampling loop without changing planning time."],"supporting_citations":[{"why":"Supplies the random multipartite graph tensor discretization structure that MTP adapts from motion planning to control sampling.","marker":"[Le et al., 2025]"},{"why":"Defines Predictive Sampling, the local MPC baseline and the source of the nominal-control initialization and first-control execution convention used in the algorithm.","marker":"[Howell et al., 2022]"},{"why":"Defines MPPI, the local baseline whose softmax weighted update is adapted for the modified Cross-Entropy update in MTP.","marker":"[Williams et al., 2017]"},{"why":"Defines OpenAI-ES, the evolutionary baseline used for comparison and the antithetic-sampling scheme.","marker":"[Salimans et al., 2017]"},{"why":"Defines Diffusion Evolution, the evolutionary baseline used for comparison that the paper claims to outperform.","marker":"[Zhang et al., 2024]"},{"why":"Supplies the recursive B-spline basis construction used to interpolate tensor-sampled waypoints into smooth control trajectories.","marker":"[de Boor, 1973]"},{"why":"Supplies the Akima spline interpolation method used as the second smooth interpolation option for control trajectories.","marker":"[Akima, 1974]"}],"fun_headline_variants":["Tensor sampling expands MPC exploration space","Random graph tensors produce high-entropy MPC paths","MPC scales exploration via tensor graph sampling","Tensor path entropy avoids local-minima traps in MPC","Asymptotic coverage for MPC via tensor paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The asymptotic path-coverage guarantee is proven only for linear interpolation in the infinite limit, but the paper's reported success relies on B-spline and Akima interpolation on very small graphs ($M=2$ to $3$), so the argument assumes the linear denseness result transfers to spline-interpolated paths at practical sizes; that transfer is asserted rather than proved.","fun_headline_variants_meta":{"raw":{"variants":["Tensor sampling expands MPC exploration space","Random graph tensors produce high-entropy MPC paths","MPC scales exploration via tensor graph sampling","Tensor path entropy avoids local-minima traps in MPC","Asymptotic coverage for MPC via tensor paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001101,"raw_usage":{"total_tokens":4651,"prompt_tokens":1062,"completion_tokens":3589,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":3520}},"tokens_in":678,"tokens_out":3589,"duration_ms":24444,"temperature":1.0,"reasoning_tokens":3520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:28:19.295248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a control task with a narrow cost basin that can only be reached by a specific smooth trajectory whose waypoint sequence is not captured by the small $M,N$ grid. If MTP with $M=2-3$, $N=50-100$ and B-spline or Akima interpolation cannot concentrate samples there while a linear-interpolation variant or a denser graph can, the coverage claim for practical settings fails. More directly, compute the distance from a family of spline-interpolated paths to the nearest linear-interpolated graph path for the $M,N$ values used in Table 3; if the distance does not shrink toward zero or if the spline paths fail to intersect the low-cost region, the transfer assumption is false.","supporting_citations":[{"cited_title":"Package for calculating with b-splines","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive B-spline basis construction used to interpolate tensor-sampled waypoints into smooth control trajectories."},{"cited_title":"A method of bivariate interpolation and smooth surface fitting based on local procedures","cited_arxiv_id":null,"evidence_quote":"Supplies the Akima spline interpolation method used as the second smooth interpolation option for control trajectories."}],"review_version":1}