Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Cross-Entropy Optimization of Physically Grounded Task and Motion Plans

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper claims that contact-rich task-and-motion plans can be found and directly executed by optimizing controller parameters inside a GPU-parallel physics simulator, without modeling contacts explicitly.

desk verdict A solid, honest integration of CE optimization with parallel physics simulation for TAMP; evidence is thin but the method is clear and the limitations are stated up front. read the letter →

arxiv 2512.11571 v2 pith:GA5QKALQ submitted 2025-12-12 cs.RO

Andreu Matoses Gimenez , Nils Wilde , Chris Pek , Javier Alonso-Mora This is my paper
classification cs.RO
keywords taskandmotionplanningcross-entropyoptimizationphysicssimulationcontact-richmanipulationmobilemanipulatorplanrealizationGPUparallelizationPDDL
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 addresses the gap between high-level task plans and physically reliable execution. It proposes to generate a candidate symbolic plan and then search over the continuous parameters of the robot's motion controllers using cross-entropy optimization, where every candidate parameter set is rolled out in a GPU-parallel physics simulator that captures dynamics and contacts. Because the optimizer reuses the same controllers that run on the real robot, the resulting plan can be executed as-is. The authors demonstrate the method on a mobile manipulator that exploits a ramp to slide an object onto a table and on a push task, and they show that the resulting behaviors would be hard to obtain with conventional samplers that require explicit contact modeling. A careful reader would care because this offers a route to planning behaviors that are literally executable while still accounting for complex contact physics.

What carries the argument

The load-bearing mechanism is the cross-entropy optimizer over controller parameters, coupled with a batched physics simulator. Each action in the symbolic plan is tied to a motion controller with continuous parameters (approach positions, grasp orientations, release poses). The optimizer maintains a factorized normal distribution over these parameters; at each iteration it samples thousands of full-plan parameter vectors, simulates them in parallel with contact dynamics, keeps the feasible elite samples with lowest cost, and updates the distribution to match them. The symbolic layer only filters logically impossible action sequences; all physical feasibility is deferred to the simulation. T

What would settle it

Execute the best plan found for the ramp setup on hardware twenty times; if the success rate drops well below 100% due to contact variability in the slide, the claim that the same controllers can 'directly execute the computed plans' would be weakened. Alternatively, deliberately choose initial sampling regions that exclude all feasible solutions in simulation and show the method cannot recover, confirming the critical dependence on initial regions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a plan's feasibility and cost do not have to be captured by hand-crafted geometric constraints; they can emerge from simulation. Given a symbolic plan, the method assigns a probability distribution to each action controller parameter, draws batches of hundreds of parameter vectors, simulates the whole plan for each under full rigid-body dynamics with contacts, and uses the feasible lowest-cost samples to reshape the distributions via cross-entropy. The authors report that after a few iterations the sampler concentrates on low-cost realizations, such as letting the cube slide down a ramp, and that the same controller parameters execute on th

Load-bearing premise

The pipeline can only start if the user-supplied initial sampling regions (the rings around tables, the grasp-orientation box, the release-pose cuboid) contain at least one parameter vector for which every action succeeds; otherwise the cross-entropy update has no feasible samples to learn from, as the paper's limitation section acknowledges.

Editorial extensions

If this is right

  • Plans that exploit environmental contacts, such as sliding an object down a ramp or pushing a block into a box, can be found automatically rather than hand-coded.
  • Because the optimized action parameters are inputs to the same controllers used on hardware, a realized plan can be executed without a separate trajectory-repair step.
  • The approach can produce multiple distinct solution modes (e.g., ramp, left bypass, right bypass) and, with broad initial sampling, tends to converge to the lowest-cost mode.
  • Feasibility rates rise sharply over cross-entropy iterations, from roughly 0.1% initially to a majority of samples near convergence.
  • The method inherits the cost of the simulator: each iteration initially takes minutes, though it drops as samples concentrate.

Reading between the lines

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

  • Because the method only searches parameters of pre-designed controllers, it cannot discover action sequences or controller behaviors outside the provided action schemas; a natural extension would be to also optimize over the discrete choice of controllers, which the paper leaves fixed.
  • The reported 35 s vs 21 s execution-time gap suggests the simulator underestimates actuator limits and ground friction; a testable extension is to calibrate the simulator's friction and damping parameters on the real robot to close the gap.
  • The dependence on an initial sampling region that contains a feasible parameter vector means the method is only as good as the domain knowledge used to pick those regions; one could test robustness by shrinking or shifting the rings and cuboids until the method fails.
  • If the same pipeline were applied to tasks with multiple interacting objects, the factorized normal distribution could struggle with multi-modal or disconnected solution sets; modeling the parameter distribution as a mixture would be a direct next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a TAMP method that separates symbolic task planning from continuous plan realization. A PDDL planner produces a symbolic action sequence; a cross-entropy (CE) algorithm then optimizes the parameters of low-level controllers in a GPU-parallelized physics simulator (Isaac Gym), so that dynamics, contacts, and controller behavior are evaluated directly. The method is demonstrated on two tasks: a pick-and-place task with a ramp or box obstacle and a move-and-push task, in simulation (10 seeds each) and in a single real-robot run for each task. The main claim is that this approach finds low-cost, executable plan realizations that exploit contacts and environment geometry without explicitly modeling these effects in the symbolic layer.

Significance. If the claims hold, the paper makes a useful contribution to contact-rich TAMP: it replaces hand-modeled feasibility checks with massively parallel simulation, handles multi-modal and disconnected solution regions through CE sampling, and reuses the same controllers in simulation and on hardware, reducing the sim-to-real gap. The approach is simple, reproducible, and the authors provide code and a website. The qualitative demonstrations of ramp-sliding and stick-pushing are compelling evidence that the method can discover contact-based solutions that are hard to encode in classical TAMP samplers. However, the paper's current evidence is narrow: two scenarios, ten seeds, no baseline comparison, no error bars, and only one real-robot run per task. The central feasibility mechanism also depends critically on hand-specified initial sampling regions, a limitation the authors explicitly acknowledge but do not address.

major comments (3)
  1. [Section IV] The empirical evaluation is thin relative to the paper's central claims. Only two tasks are considered, each with 10 random seeds, and Figure 5 reports no error bars, confidence intervals, or per-seed spread. There is also no baseline comparison with existing TAMP realization methods, despite the claim in Section V that behavior from contacts 'could not be modeled in previous sampling-based methods [3], [4]'. To support the generality claim, the paper should include quantitative comparisons (e.g., success rate, cost, and compute time against a sampling-based TAMP solver or an ablation without simulation), and report variance across seeds and across real-robot repetitions. Without this, the demonstration remains proof-of-concept rather than a validated method.
  2. [Section IV-C] The real-world validation is a single run per task. The reported pick-and-place execution takes 35 s versus 21 s in simulation, a 67% discrepancy attributed to actuator limitations and friction; the push task has a 3 s difference. The paper states that the sim-to-real gap is small, but one run without repeated trials or quantitative success criteria does not establish reliable direct execution. For the central claim that the same controllers transfer directly to the real system, at least multiple repetitions, success/failure counts, and a discussion of how the 14 s discrepancy affects plan feasibility or cost should be provided. At minimum, the paper should soften the 'directly execute' claim to reflect the observed fidelity limits.
  3. [Algorithm 1 and Section III-B] The CE algorithm is underspecified in ways that affect reproducibility. The convergence criterion in line 7 ('while θ not converged') is never defined; the artificial noise added when only one feasible sample is available is not described; and the update rule for μ and σ is only stated informally ('the resulting μ and σ of the elite parameter samples are used'). Since the paper's contribution is algorithmic, these details should be specified precisely, including the stopping rule used in the reported experiments.
minor comments (4)
  1. [Throughout] Typos and formatting issues: 'untractable' should be 'intractable', 'task planing' should be 'task planning', 'a priory' should be 'a priori', and 'T abeSurf ace' in Section IV-A is a broken word. Reference [21] and [27] are the same Fast Downward paper and should be merged.
  2. [Section IV-A] The initial sampling regions are described informally. For example, the quaternion is 'initially sampled uniformly within a 0.3 size hyper-rectangle'—this is ambiguous: is it an L∞ box of side length 0.3 in R^4? Similarly, the 20 cm padding used by A* appears only in the caption of Figure 4 and should be stated in the setup.
  3. [Section IV-A and Figure 5] The cost functions Jm, Jg, Jp are defined as 1 per time step until the action succeeds, making the total cost effectively a completion time. This should be stated explicitly, and the handling of failed actions (e.g., whether the horizon is truncated) should be clarified. Figures 5a-c would benefit from shaded standard deviation bands or individual seed trajectories.
  4. [Section V-A] The limitation paragraph says 'for real-time execution, one can use the approximated distributions ... as long as the task and environment do not change.' This is a useful observation, but it is not demonstrated; consider adding a short experiment or at least a quantitative estimate of how much compute is saved in subsequent runs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the CE optimization evaluates controller parameters against independent simulator costs and success predicates; the only author-overlap citation (fabrics controller) is an engineering component validated on the real robot, not a load-bearing logical premise.

full rationale

The derivation chain is self-contained as an optimization pipeline. A PDDL planner outputs a candidate symbolic plan; Algorithm 1 samples controller parameters from initial uniform regions; IsaacGym rolls out each full plan under the dynamics f; per-action success predicates (e.g., ee_in(z): ||z - p_ee|| <= epsilon) and the terminal goal psi_g(x_T) are Boolean conditions that do not depend on the optimized values in a way that would force the result; and the CE update uses the top N_e feasible, lowest-cost samples to update the sampling distribution. The cost functions are per-time-step penalties, not functions of the fitted distribution parameters. Thus the optimized parameters are not, by construction, equal to their inputs. The paper explicitly flags its main limitation: 'If the initially chosen parameter regions are too wide or if the task has a small solution space, we may find that none of the initial candidate plans is feasible. Currently, we use domain knowledge to provide the algorithm with a good set of initial parameter regions' (Section V-A). This is a real precondition and a robustness limitation, but it is not circular: the paper does not derive the initial regions from the reported solutions, and it acknowledges the dependence. Similarly, the sim-to-real gap ('the realized plan in the real system takes 35 s to execute, while the simulated one takes 21 s', Section IV-C) is a fidelity limitation, not a renamed prediction. The only author-overlap citation is [24] for geometric fabrics, used as the whole-body controller. The method's central claim does not reduce to that citation: the fabrics controller is a component whose behavior is evaluated in the paper's own simulator and real-robot experiments. No uniqueness theorem, ansatz, or fitted quantity is imported from the authors' prior work, and no known result is renamed as a new contribution. Therefore the appropriate finding is no significant circularity.

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

No new physical entities are introduced. The central claim rests on two load-bearing domain assumptions—simulator fidelity and initial sampling feasibility—plus standard CE convergence behavior. Free parameters are hand-set task and algorithm choices.

free parameters (3)
  • Initial sampling region geometries = Table1 ring r=1.1/R=1.5 m; Table2 ring r=1.2/R=1.7 m; Cube quaternion hyper-rect side 0.3; Target translation hyper-rect
    Hand-chosen from domain knowledge; the paper states initial regions are crucial and poor choices can make all initial samples infeasible.
  • CE algorithm settings = n_envs=3000 decreasing linearly to 300; N_e=50
    Fixed without sensitivity analysis; they affect exploration/exploitation balance and compute time.
  • Action cost and success tolerances = J=1 per timestep until success; epsilon tolerances unspecified
    The 'low-cost' objective and feasibility thresholds are defined by these choices; no ablation or sensitivity is reported.
assumptions (3)
  • domain assumption The IsaacGym physics simulator with identical controllers accurately models the real system's dynamics and contacts.
    Central to sim-to-real transfer; the paper itself reports a 14 s discrepancy in pick-and-place execution time, indicating imperfect fidelity. Invoked throughout Section IV.
  • domain assumption The initial uniform sampling distributions over parameter regions contain at least one feasible plan realization.
    Algorithm 1 cannot update without feasible samples; the paper states 'we may find that none of the initial candidate plans is feasible' in Section V-A.
  • standard math Cross-entropy optimization with independent normal distributions converges to low-cost regions of a potentially multimodal parameter space.
    Standard CE method [22] assumptions; the paper acknowledges modes may be overrepresented and distribution collapses to a single local minimum (Section IV-B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Entropy Optimization of Physically Grounded Task and Motion Plans." pith.science (2026). https://pith.science/paper/GA5QKALQ

@misc{pith2026251211571,
  author       = {Pith},
  title        = {Pith review of: Cross-Entropy Optimization of Physically Grounded Task and Motion Plans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GA5QKALQ}},
  note         = {Machine review of arXiv:2512.11571}
}
read the original abstract

Autonomously performing tasks often requires robots to plan high-level discrete actions and continuous low-level motions to realize them. Previous TAMP algorithms have focused mainly on computational performance, completeness, or optimality by making the problem tractable through simplifications and abstractions. However, this comes at the cost of the resulting plans potentially failing to account for the dynamics or complex contacts necessary to reliably perform the task when object manipulation is required. Additionally, approaches that ignore effects of the low-level controllers may not obtain optimal or feasible plan realizations for the real system. We investigate the use of a GPU-parallelized physics simulator to compute realizations of plans with motion controllers, explicitly accounting for dynamics, and considering contacts with the environment. Using cross-entropy optimization, we sample the parameters of the controllers, or actions, to obtain low-cost solutions. Since our approach uses the same controllers as the real system, the robot can directly execute the computed plans. We demonstrate our approach for a set of tasks where the robot is able to exploit the environment's geometry to move an object. Website and code: https://andreumatoses.github.io/research/parallel-realization

Figures

Figures reproduced from arXiv: 2512.11571 by the authors.

Figure 1
Figure 1. (a) Mobile manipulator performing a sequential pick and place [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our TAMP framework. (a) The scenario specification involves defining symbolic actions the robot can perform, e.g., moving to [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Simulation results of: a) successful plans for the setup with a ramp-shaped obstacle, and b), the setup without ramp. When the ramp is present, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Plans simulated from the sampled parameters for different iterations of the CE optimization. The trajectories displayed in orange tones show the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Top: Evolution of the cost of the best-sampled plan realization for each problem scenario. Bottom: Change in percentage of sampled plans that [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Plan executed in the real system, using the same low level [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 1 linked inside Pith

  1. [3]

    PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning,

    C. R. Garrett, T. Lozano-P ´erez, and L. P. Kaelbling, “PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning,”Proceedings of the International Conference on Automated Planning and Scheduling, vol. 30, pp. 440–448, Jun. 2020

    2020

  2. [4]

    Task and Motion Informed Trees (TMIT*): Almost-Surely Asymptotically Optimal In- tegrated Task and Motion Planning,

    W. Thomason, M. P. Strub, and J. D. Gammell, “Task and Motion Informed Trees (TMIT*): Almost-Surely Asymptotically Optimal In- tegrated Task and Motion Planning,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11 370–11 377, Oct. 2022

    2022

  3. [1]

    Logic-geometric programming: an optimization-based approach to combined task and motion planning,

    M. Toussaint, “Logic-geometric programming: an optimization-based approach to combined task and motion planning,” inProceedings of the 24th International Conference on Artificial Intelligence, ser. IJCAI’15. Buenos Aires, Argentina: AAAI Press, Jul. 2015, pp. 1930–1936

    2015

  4. [2]

    Integrated Task and Motion Planning,

    C. R. Garrett, R. Chitnis, R. Holladay, B. Kim, T. Silver, L. P. Kael- bling, and T. Lozano-P ´erez, “Integrated Task and Motion Planning,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 4, no. V olume 4, 2021, Oct. 2020

    2021

  5. [5]

    Differentiable Physics and Stable Modes for Tool-Use and Manipulation Planning,

    M. Toussaint, K. Allen, K. Smith, and J. Tenenbaum, “Differentiable Physics and Stable Modes for Tool-Use and Manipulation Planning,” inRobotics: Science and Systems XIV, vol. 14, Jun. 2018

    2018

  6. [6]

    Isaac gym: High performance gpu-based physics simulation for robot learning,

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Mack- lin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac gym: High performance gpu-based physics simulation for robot learning,” 2021

    2021

  7. [7]

    Forge: Force-guided exploration for robust contact-rich manipulation under uncertainty,

    M. Noseworthy, B. Tang, B. Wen, A. Handa, N. Roy, D. Fox, F. Ramos, Y . Narang, and I. Akinola, “Forge: Force-guided exploration for robust contact-rich manipulation under uncertainty,”arXiv preprint arXiv:2408.04587, 2024

    arXiv 2024

  8. [8]

    Accelerated policy learning with parallel differen- tiable simulation,

    J. Xu, V . Makoviychuk, Y . Narang, F. Ramos, W. Matusik, A. Garg, and M. Macklin, “Accelerated policy learning with parallel differen- tiable simulation,” inInternational Conference on Learning Represen- tations, 2021

    2021

Show all 27 references
  1. [9]

    PDDL-the planning domain definition language,

    D. McDermott, M. Ghallab, A. Howe, C. A. Knoblock, A. Ram, M. Veloso, D. S. Weld, and D. Wilkins, “PDDL-the planning domain definition language,” 1998

    1998

  2. [10]

    A Survey of Optimization-based Task and Motion Planning: From Classical To Learning Approaches,

    Z. Zhao, S. Cheng, Y . Ding, Z. Zhou, S. Zhang, D. Xu, and Y . Zhao, “A Survey of Optimization-based Task and Motion Planning: From Classical To Learning Approaches,”IEEE/ASME Transactions on Mechatronics, pp. 1–27, 2024

    2024

  3. [11]

    An incremental constraint-based framework for task and motion planning,

    N. T. Dantam, Z. K. Kingston, S. Chaudhuri, and L. E. Kavraki, “An incremental constraint-based framework for task and motion planning,” The International Journal of Robotics Research, vol. 37, no. 10, pp. 1134–1151, Sep. 2018

    2018

  4. [12]

    Multi-bound tree search for logic- geometric programming in cooperative manipulation domains,

    M. Toussaint and M. Lopes, “Multi-bound tree search for logic- geometric programming in cooperative manipulation domains,” in 2017 IEEE International Conference on Robotics and Automation (ICRA), May 2017, pp. 4044–4051

    2017

  5. [13]

    Newton methods for k-order Markov Constrained Motion Problems,

    M. Toussaint, “Newton methods for k-order Markov Constrained Motion Problems,” Jul. 2014

    2014

  6. [14]

    Force-and-Motion Constrained Planning for Tool Use,

    R. Holladay, T. Lozano-Perez, and A. Rodriguez, “Force-and-Motion Constrained Planning for Tool Use,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Macau, China: IEEE, Nov. 2019, pp. 7409–7416. 9

    2019

  7. [15]

    Sampling-based methods for factored task and motion planning,

    C. R. Garrett, T. Lozano-P ´erez, and L. P. Kaelbling, “Sampling-based methods for factored task and motion planning,”The International Journal of Robotics Research, vol. 37, no. 13-14, pp. 1796–1825, Dec. 2018

    2018

  8. [16]

    NLP Sampling: Com- bining MCMC and NLP Methods for Diverse Constrained Sampling,

    M. Toussaint, C. V . Braun, and J. Ortiz-Haro, “NLP Sampling: Com- bining MCMC and NLP Methods for Diverse Constrained Sampling,” Jul. 2024

    2024

  9. [17]

    Stein Variational Evolution Strategies,

    C. V . Braun, R. T. Lange, and M. Toussaint, “Stein Variational Evolution Strategies,” Oct. 2024

    2024

  10. [18]

    Sequence- of-Constraints MPC: Reactive Timing-Optimal Control of Sequential Manipulation,

    M. Toussaint, J. Harris, J.-S. Ha, D. Driess, and W. H ¨onig, “Sequence- of-Constraints MPC: Reactive Timing-Optimal Control of Sequential Manipulation,” in2022 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), Sep. 2022

    2022

  11. [19]

    Task and Motion Planning for Execution in the Real,

    T. Pan, R. Shome, and L. E. Kavraki, “Task and Motion Planning for Execution in the Real,”IEEE Transactions on Robotics, vol. 40, pp. 3356–3371, Jun. 2024

    2024

  12. [20]

    Partially Observable Task and Motion Planning with Uncertainty and Risk Awareness,

    A. Curtis, G. Matheos, N. Gothoskar, V . Mansinghka, J. Tenenbaum, T. Lozano-P´erez, and L. P. Kaelbling, “Partially Observable Task and Motion Planning with Uncertainty and Risk Awareness,” inRobotics: Science and Systems XX. arXiv, Mar. 2024

    2024

  13. [21]

    The Fast Downward Planning System,

    M. Helmert, “The Fast Downward Planning System,”Journal of Artificial Intelligence Research, vol. 26, pp. 191–246, Jul. 2006

    2006

  14. [22]

    Chapter 3 - The Cross-Entropy Method for Optimization,

    Z. I. Botev, D. P. Kroese, R. Y . Rubinstein, and P. L’Ecuyer, “Chapter 3 - The Cross-Entropy Method for Optimization,” inHandbook of Statistics, ser. Handbook of Statistics, C. R. Rao and V . Govindaraju, Eds. Elsevier, Jan. 2013, vol. 31, pp. 35–59

    2013

  15. [23]

    Optimization Fabrics,

    N. D. Ratliff, K. Van Wyk, M. Xie, A. Li, and M. A. Rana, “Optimization Fabrics,” Aug. 2020

    2020

  16. [24]

    Dynamic Optimization Fabrics for Motion Generation,

    M. Spahn, M. Wisse, and J. Alonso-Mora, “Dynamic Optimization Fabrics for Motion Generation,”IEEE Transactions on Robotics, vol. 39, pp. 2684–2699, Mar. 2023

    2023

  17. [25]

    Rie- mannian motion policies,

    N. D. Ratliff, J. Issac, D. Kappler, S. Birchfield, and D. Fox, “Rie- mannian motion policies,” 2018

    2018

  18. [26]

    A formal basis for the heuristic determination of minimum cost paths,

    P. Hart, N. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,”IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968

    1968

  19. [27]

    The fast downward planning system,

    M. Helmert, “The fast downward planning system,”Journal of Artifi- cial Intelligence Research, vol. 26, p. 191–246, Jul. 2006

    2006

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.