Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

CSC-MPPI: A Novel Constrained MPPI Framework with DBSCAN for Reliable Obstacle Avoidance

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Clustering-based MPPI reports zero collisions in obstacle tests

desk verdict Reasonable engineering idea with an unsupported guarantee: the primal-dual step ignores obstacle constraints and cluster averages are not proved feasible. read the letter →

arxiv 2506.16386 v2 pith:ZFIQXNI6 submitted 2025-06-19 cs.RO

Leesai Park , Keunwoo Jang , Sanghyun Kim This is my paper · ORCID
classification cs.RO
keywords ModelPredictivePathIntegralcontrolMPPIDBSCANprimal-dualgradientmethodhardconstraintsobstacleavoidancetrajectoryclusteringmobilerobotnavigation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes CSC-MPPI, a version of Model Predictive Path Integral (MPPI) control—a sampling-based controller that averages many stochastic rollouts—and claims it enforces hard state and input constraints rather than soft penalties. It inserts a primal-dual gradient step (an iterative update that alternates between the control variable and penalty multipliers) to shift every sampled control sequence into the feasible region, then groups the feasible samples with DBSCAN and executes the lowest-cost cluster's weighted average instead of the global average. The motivation is concrete: the global averaging step in ordinary MPPI can combine individually safe samples into an unsafe trajectory, and the paper's experiments report that clustering removes those violations. If the claim holds, sampling-based MPC gains a way to respect strict constraints while keeping the sampling paradigm's flexibility.

What carries the argument

The load-bearing object is the constrained sampling-and-selection loop in Algorithm 1. Sampled deviations $\delta^k u$ are moved toward feasibility by alternating the primal update $v_t \leftarrow v_t - \alpha \circ \nabla_{v_t} \mathcal{L}_t$ with projected dual updates $\mu_t \leftarrow \max(0, \mu_t + \beta_1 \circ (v - v_t))$ until the KKT conditions hold; DBSCAN then clusters the pairs $D = \{(\delta U^k, S^k)\}$ by control deviation and cost, and a softmax-weighted average is taken within each cluster, with the lowest-cost cluster supplying the executed control. This machinery is the bridge that converts probabilistic sampling into a procedure that claims to execute only feasible representatives.

What would settle it

Construct an obstacle field where the start and goal are separated by an obstacle and the feasible samples split evenly between two routes around it, then run CSC-MPPI repeatedly with different random seeds and check whether the elected cluster's average ever crosses the obstacle; the paper's own Environment 2 comparison provides the template, since the no-DBSCAN variant already violated constraints in 20% of runs.

Watch

Extended reading notes

Core claim

The central claim is that constraint satisfaction in MPPI can be made hard, not soft, by changing what happens between sampling and execution. After drawing control sequences from a Gaussian proposal, the algorithm alternates a gradient step on the control and projected updates on Lagrange multipliers until the KKT conditions are satisfied, which shifts each sample into the feasible region; it then clusters the feasible samples by control deviation and cost with DBSCAN, computes a softmax-weighted average inside each cluster, and selects the lowest-cost cluster average as the action. The paper reports that this pipeline achieved a 0% collision rate across its simulations and real-world runs, and that in its second environment the version without DBSCAN satisfied constraints only 80% of the time while the full version did so 100% of the time, attributing the difference to clustering rather than to projection alone.

Load-bearing premise

The guarantee rests on the unproven premise that the cost-weighted average of the feasible control sequences inside every DBSCAN cluster is itself feasible, because that average is what the robot executes.

Editorial extensions

If this is right

  • With only $K=20$ samples, CSC-MPPI kept a 0% collision rate in the dynamic-obstacle environment where standard MPPI collided in 80% of runs.
  • The clustering step is the part that prevents averaging-induced infeasibility: in the second environment the no-DBSCAN variant satisfied constraints in 80% of runs, while the full method satisfied them in 100%.
  • Computation remains real-time compatible, about 6–9 ms per iteration on GPU in simulation and about 6.6 ms average on CPU in hardware, including clustering.
  • In ten real-world runs with seven static obstacles, the method again produced zero collisions.
  • Because DBSCAN discards outliers before averaging, high-cost or geometrically inconsistent samples cannot drag the chosen control toward a constraint-violating direction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the two-step pattern holds up beyond these testbeds, it suggests a general repair for sampling-based predictive controllers that suffer from the failure mode in which the average of two safe plans is unsafe: project samples to feasibility first, then cluster before averaging.
  • A direct next experiment is a symmetric environment where start and goal sit on opposite sides of an obstacle and feasible samples split evenly between left and right routes; the method's behavior there would reveal whether DBSCAN reliably chooses one homotopy class instead of blending them.
  • The hardware runs used input clamping in place of the full gradient projection for speed, so the natural extension is a comparison of clamping versus projection on coupled constraints such as joint limits plus torque limits, where clamping is expected to degrade.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes CSC-MPPI, a constrained Model Predictive Path Integral framework that first applies a primal-dual gradient update to shift randomly sampled control sequences toward a feasible region defined by state and control constraints, and then uses DBSCAN to cluster the adjusted trajectories and select the lowest-cost cluster representative as the executed control input. The authors claim that this pipeline guarantees hard constraint satisfaction for both states and control inputs, and they support the claim with simulations (two environments), an ablation that removes the DBSCAN step, and real-world experiments on a LIMO robot, all reporting zero collisions and improved path lengths compared with standard MPPI.

Significance. If the hard-constraint guarantee were established, the paper would make a meaningful contribution to sampling-based predictive control, where soft penalties or probabilistic barrier functions are currently the norm. The paper is clearly structured, provides complete pseudocode, and includes both a controlled ablation and hardware trials, which is commendable. However, the central contribution—the guarantee—is not proven; the paper never specifies the conditions under which the guarantee holds, and the two mechanisms it relies on (primal-dual adjustment and cluster averaging) do not, as analyzed, confer a certificate of feasibility on the executed input. The empirical results are promising (0% collision across 40 simulation and 10 hardware trials, and 100% vs 80% constraint satisfaction in the DBSCAN ablation), but 10 or 20 trials per condition are too few to substitute for a formal guarantee, and no comparison with existing constrained MPPI baselines is made. In short, the paper shows an empirically useful heuristic, not the guaranteed method promised in the title and abstract.

major comments (4)
  1. [Section III-A, Eq. (10)] The Lagrangian in Eq. (10) omits the obstacle constraint g(x_{t+1}) <= 0 as a constraint. In Eq. (9) the obstacle function appears only in the objective through the term g(kx_{t+1}) * 1{g(kX)>0}, which is a penalty for infeasible samples, while the Lagrange multipliers are associated only with the control-bounds constraints. Consequently, the paper's claim that the primal-dual iteration enforces the KKT conditions and guarantees convergence to a state-feasible solution is not justified. No convergence proof is given for the nonconvex obstacle-avoidance set, the indicator function makes the objective nonsmooth, and the referenced result [14] applies to convex-concave saddle-point problems, not to this nonconvex setting. This unproven step is the foundation of the 'hard constraint guarantee' in the Abstract and Section I.B.
  2. [Section III-B, Algorithm 1, lines 24-35] The final control input is a cost-weighted average of control sequences within a DBSCAN cluster (U* = U + sum_i w_i deltaU_i). The paper asserts that this average is feasible, but no proof is provided. The obstacle constraint g(x) = r^2 - ||x - x_obs||^2 (Eq. (18)) defines a nonconvex feasible set, and a weighted average of individually feasible control sequences can produce an infeasible state trajectory; the paper's own Fig. 5a demonstrates this failure for the no-DBSCAN variant. DBSCAN's density-based filtering does not make the feasible set convex, and if every cluster representative is infeasible, the argmin in Line 27 cannot restore feasibility. The reported 100% satisfaction for CSC-MPPI in Environment #2 is an empirical observation over 10 trials, not a guarantee. This is the load-bearing assumption of the algorithm's safety claim and it is unsupported.
  3. [Section IV.D] The real-world experiments do not exercise the proposed constraint-handling mechanism. The text after Fig. 6 states that 'we employed a clamping technique in (9) to reduce computation time,' which replaces the primal-dual gradient adjustment with direct clipping of control inputs. Clamping can enforce box constraints on the control, but it cannot enforce the obstacle state constraint g(x) <= 0, and the authors themselves note that clamping is only sufficient for 'simple constraints.' Since the paper does not report whether the obstacle constraint was active in the hardware trials, the claim that CSC-MPPI's hard-constraint enforcement is 'validated in real-world scenarios' is not supported by the reported experiment.
  4. [Section I.A and Section IV.C] The paper claims superiority over prior constrained MPPI methods (o-MPPI [10], CCS-MPPI [11], Shield-MPPI [12], SCBF-MPPI [13]) and asserts in Section I.B that this is 'the first MPPI-based approach to guarantee hard constraint satisfaction in both domains.' However, the experimental evaluation compares only against standard MPPI (Table II) and a no-DBSCAN ablation; none of the existing constrained MPPI baselines is compared. Without a formal proof of the guarantee, the novelty claim is unsubstantiated, and the empirical comparison is insufficient to differentiate CSC-MPPI from prior constrained-MPPI methods that also report improved constraint satisfaction.
minor comments (7)
  1. [Section II.B, Eqs. (7)-(8)] The variables \hat{u}_t and \tilde{u}_t appear in the importance-sampling weight formulas but are not defined in Table I or in the text, which makes the derivation of the weighted update difficult to follow.
  2. [Section III.A, Eq. (9)] The control-bounds line 'v <= kvt <= v' uses the same symbol v for both lower and upper bounds; use distinct notation such as v_min and v_max for clarity.
  3. [Section III.A, Eq. (10)] The indicator 1{g(kX)>0} is written in terms of the trajectory kX, but g is defined as a pointwise function of state; the paper should clarify whether the indicator uses the maximum, minimum, or sum of g over the trajectory.
  4. [Algorithm 1, line 8] The while loop 'while KKT Conditions not Satisfied' has no tolerance, maximum iteration count, or concrete termination criterion, making the algorithm nondeterministic and not directly reproducible.
  5. [Section III.B, DBSCAN input] The input to DBSCAN is D = {(deltaU_k, S_k)}, where deltaU_k is a control-sequence vector and S_k is a scalar cost; no distance metric or normalization between these heterogeneous components is specified, so the clustering step is not reproducible from the paper alone.
  6. [Section IV.C] The no-DBSCAN variant is reported as having a 0% collision rate but only an 80% constraint-satisfaction rate for the optimal input sequence; the paper does not explain how an input sequence that violates the obstacle constraint can avoid collision, which makes the distinction between 'collision rate' and 'constraint satisfaction rate' confusing.
  7. [Section IV.C, Fig. 5b] The caption of Fig. 5b says that colored lines represent clustered trajectories, but the representative trajectories that are the candidates for the final control are not visually distinguished; marking them would help the reader verify the claimed cluster-average selection.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the guarantee is asserted rather than proven, but no prediction or result reduces to a fit or to a load-bearing self-citation.

full rationale

The derivation chain in CSC-MPPI is not circular. The paper's central claim—that the method guarantees constraint satisfaction—is supported by two algorithmic mechanisms: a primal-dual gradient projection that iterates until KKT conditions are satisfied, and DBSCAN clustering intended to avoid the infeasibility of convex combinations of feasible controls. Neither mechanism defines its output in terms of the target claim. In Algorithm 1, each cluster representative is computed as U + sum_i w_i deltaU_i (lines 24-27) and the lowest-cost representative is selected; the paper does not define 'feasible representative' as the selected candidate, so the guarantee is not true by construction. The actual weakness is that the paper never proves that a cost-weighted average of feasible sampled controls remains feasible under the nonconvex obstacle constraint (18), and its own Fig. 5a shows the unclustered weighted average can violate constraints. That is a soundness gap, not circularity. There is also no fitted parameter relabeled as a prediction: the temperature lambda is set per environment (0.01 in Env1, 0.7 in Env2) as an experimental choice, and the reported 0% collision and 100% satisfaction rates are empirical outcomes from 20 and 10 trials, not outputs forced by the tuning. The only self-citation is reference [2], where co-author S. Kim appears; it is cited as background for traditional MPC and is not load-bearing for the MPPI contribution. No uniqueness theorem from the authors is imported, and no ansatz is smuggled in via citation. Therefore the central claim has independent algorithmic content, even if the 'guarantee' wording is stronger than what the analysis establishes.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim relies on two unproven assumptions (convergence of a nonconvex primal-dual update and feasibility of cluster averages), plus several unstated hyperparameters. These are the main load-bearing uncertainties.

free parameters (4)
  • Temperature λ = 0.01 (Env1), 0.7 (Env2)
    Chosen per environment; the paper says the higher value in Env2 was 'deliberately chosen to highlight the effect of DBSCAN.' This tuning affects the weight distribution and the reported satisfaction rates.
  • Primal-dual step sizes α, β1, β2 = not reported
    Control the convergence of the constraint-enforcement loop; values are not given, so the experiments are not reproducible without them.
  • DBSCAN parameters (eps, min_samples) = not reported
    Determine clustering and outlier rejection; not specified, although the cluster output determines the selected control input.
  • Cost weight matrices Q and H = Q=diag(10,10,0), H=diag(50,50,50)
    Hand-selected to balance goal-reaching and smoothness; these weights shape the path-length results.
assumptions (3)
  • ad hoc to paper The primal-dual gradient method converges to a feasible point for the nonconvex obstacle constraint defined by g(x) <= 0
    Section III.A states 'Iterations continue until all KKT conditions are satisfied,' but the Lagrangian (Eq. 10) includes only control bounds, not the obstacle constraint, and no convergence proof is provided for the nonconvex safe set.
  • ad hoc to paper The cost-weighted average of control sequences within a DBSCAN cluster remains feasible
    Assumed in Section III.B and Algorithm 1; the paper acknowledges averaging can violate constraints and relies on clustering to avoid it, but gives no guarantee.
  • domain assumption The kinematic model (Eq. 15) accurately predicts the robot's motion over the horizon
    The controller propagates samples with this model, and constraint enforcement depends on the propagated states.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSC-MPPI: A Novel Constrained MPPI Framework with DBSCAN for Reliable Obstacle Avoidance." pith.science (2026). https://pith.science/paper/ZFIQXNI6

@misc{pith2026250616386,
  author       = {Pith},
  title        = {Pith review of: CSC-MPPI: A Novel Constrained MPPI Framework with DBSCAN for Reliable Obstacle Avoidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZFIQXNI6}},
  note         = {Machine review of arXiv:2506.16386}
}
read the original abstract

This paper proposes Constrained Sampling Cluster Model Predictive Path Integral (CSC-MPPI), a novel constrained formulation of MPPI designed to enhance trajectory optimization while enforcing strict constraints on system states and control inputs. Traditional MPPI, which relies on a probabilistic sampling process, often struggles with constraint satisfaction and generates suboptimal trajectories due to the weighted averaging of sampled trajectories. To address these limitations, the proposed framework integrates a primal-dual gradient-based approach and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to steer sampled input trajectories into feasible regions while mitigating risks associated with weighted averaging. First, to ensure that sampled trajectories remain within the feasible region, the primal-dual gradient method is applied to iteratively shift sampled inputs while enforcing state and control constraints. Then, DBSCAN groups the sampled trajectories, enabling the selection of representative control inputs within each cluster. Finally, among the representative control inputs, the one with the lowest cost is chosen as the optimal action. As a result, CSC-MPPI guarantees constraint satisfaction, improves trajectory selection, and enhances robustness in complex environments. Simulation and real-world experiments demonstrate that CSC-MPPI outperforms traditional MPPI in obstacle avoidance, achieving improved reliability and efficiency. The experimental videos are available at https://cscmppi.github.io

Figures

Figures reproduced from arXiv: 2506.16386 by the authors.

Figure 1
Figure 1. Comparison of standard MPPI and CSC-MPPI. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CSC-MPPI process. (a) The gray circle represents an obstacle, while the green and red distributions [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Simulation environments used for performance eval [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Trajectory comparison between CSC-MPPI and its [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Snapshots of the real-world experiment. The red line [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Projection-Retraction MPPI: Exact Constraint-Manifold Control for Manipulators

    cs.RO 2026-08 conditional novelty 6.0 of 10

    A projection-retraction MPPI variant that enforces equality and inequality constraints inside sampled rollouts and returns commands that satisfy the closed-chain constraint to numerical tolerance.

  2. Multi-Modal Model Predictive Path Integral Control for Collision Avoidance

    cs.RO 2025-08 conditional novelty 6.0 of 10

    A multi-modal sampling-based controller, combining Sobol sequences with analytical braking, acceleration, and evasive modes, outperforms a standard MPPI baseline in simulated high- and low-friction collision-avoidance...

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages · cited by 2 Pith papers

  1. [14]

    Linear convergence of the primal-dual gradient method for convex-concave saddle point problems without strong con- vexity,

    S. S. Du and W. Hu, “Linear convergence of the primal-dual gradient method for convex-concave saddle point problems without strong con- vexity,” inThe 22nd International Conference on Artificial Intelligence and Statistics. PMLR, 2019, pp. 196–205

    work page 2019

  2. [10]

    Output-sampled model predictive path integral control (o-mppi) for increased efficiency,

    L. L. Yan and S. Devasia, “Output-sampled model predictive path integral control (o-mppi) for increased efficiency,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 14 279–14 285

    work page 2024

  3. [11]

    Constrained covariance steering based tube-mppi,

    I. M. Balci, E. Bakolas, B. Vlahov, and E. A. Theodorou, “Constrained covariance steering based tube-mppi,” in 2022 American Control Conference (ACC). IEEE, 2022, pp. 4197–4202

    work page 2022

  4. [12]

    Shield model predictive path integral: A computationally efficient robust mpc method using control barrier functions,

    J. Yin, C. Dawson, C. Fan, and P. Tsiotras, “Shield model predictive path integral: A computationally efficient robust mpc method using control barrier functions,” IEEE Robotics and Automation Letters , 2023

    work page 2023

  5. [13]

    Path integral methods with stochastic control barrier functions,

    C. Tao, H.-J. Yoon, H. Kim, N. Hovakimyan, and P. V oulgaris, “Path integral methods with stochastic control barrier functions,” in 2022 IEEE 61st Conference on Decision and Control (CDC) . IEEE, 2022, pp. 1654–1659

    work page 2022

  6. [1]

    Aggressive driving with model predictive path integral control,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Aggressive driving with model predictive path integral control,” in 2016 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2016, pp. 1433–1440

    2016

  7. [2]

    Online multi-contact receding horizon planning via value function approximation,

    J. Wang, S. Kim, T. S. Lembono, W. Du, J. Shim, S. Samadi, K. Wang, V . Ivan, S. Calinon, S. Vijayakumar et al. , “Online multi-contact receding horizon planning via value function approximation,” IEEE Transactions on Robotics , 2024

    work page 2024

  8. [3]

    Mpc for robot manipula- tors with integral sliding modes generation,

    G. P. Incremona, A. Ferrara, and L. Magni, “Mpc for robot manipula- tors with integral sliding modes generation,” IEEE/ASME Transactions on Mechatronics, vol. 22, no. 3, pp. 1299–1307, 2017

    work page 2017

Show all 19 references
  1. [4]

    Implementation of nonlinear model predictive path-following control for an industrial robot,

    T. Faulwasser, T. Weber, P. Zometa, and R. Findeisen, “Implementation of nonlinear model predictive path-following control for an industrial robot,” IEEE Transactions on Control Systems Technology , vol. 25, no. 4, pp. 1505–1511, 2016

    2016

  2. [5]

    A generalized iterative lqg method for locally- optimal feedback control of constrained nonlinear stochastic systems,

    E. Todorov and W. Li, “A generalized iterative lqg method for locally- optimal feedback control of constrained nonlinear stochastic systems,” in 2005 American Control Conference (ACC) . IEEE, 2005, pp. 300– 306

    2005

  3. [6]

    Squash-box feasibility driven differential dynamic programming,

    J. Marti-Saumell, J. Sol `a, C. Mastalli, and A. Santamaria-Navarro, “Squash-box feasibility driven differential dynamic programming,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 7637–7644

    2020

  4. [7]

    Control-limited differential dynamic programming,

    Y . Tassa, N. Mansard, and E. Todorov, “Control-limited differential dynamic programming,” in 2014 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2014, pp. 1168–1175

    2014

  5. [8]

    Differential dynamic programming with nonlinear constraints,

    Z. Xie, C. K. Liu, and K. Hauser, “Differential dynamic programming with nonlinear constraints,” in 2017 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2017, pp. 695–702

    2017

  6. [9]

    An integrated perturbation analysis and sequential quadratic programming approach for model predictive control,

    R. Ghaemi, J. Sun, and I. V . Kolmanovsky, “An integrated perturbation analysis and sequential quadratic programming approach for model predictive control,” Automatica, vol. 45, no. 10, pp. 2412–2418, 2009

    2009

  7. [15]

    Dbscan clustering algorithm based on density,

    D. Deng, “Dbscan clustering algorithm based on density,” in 2020 7th international forum on electrical engineering and automation (IFEEA). IEEE, 2020, pp. 949–953

    2020

  8. [16]

    Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,” IEEE Transactions on Robotics, vol. 34, no. 6, pp. 1603–1622, 2018

    2018

  9. [17]

    Generalized lagrange multiplier method and kkt conditions with an application to distributed optimization,

    M. Li, “Generalized lagrange multiplier method and kkt conditions with an application to distributed optimization,” IEEE Transactions on Circuits and Systems II: Express Briefs , vol. 66, no. 2, pp. 252– 256, 2018

    2018

  10. [18]

    Path integral control with rollout clustering and dynamic obstacles,

    S. Patrick and E. Bakolas, “Path integral control with rollout clustering and dynamic obstacles,” in 2024 American Control Conference (ACC). IEEE, 2024, pp. 809–814

    2024

  11. [19]

    Bic-mppi: Goal-pursuing, sampling-based bidi- rectional rollout clustering path integral for trajectory optimization,

    M. Jung and K. Kim, “Bic-mppi: Goal-pursuing, sampling-based bidi- rectional rollout clustering path integral for trajectory optimization,” arXiv preprint arXiv:2410.06493 , 2024

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.