Pith. sign in

REVIEW 2 major objections 5 minor 24 references

Optimization-based Posture Generation for Whole-body Contact Motion by Contact Point Search on the Body Surface

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that whole-body contact postures can be generated by one gradient-based optimization that simultaneously solves for joint angles and contact points on the robot's body surface.

desk verdict A practical contact-point parameterization for whole-body posture optimization, with real-robot demos; the edge-crossing gradient is approximate but the empirical support is decent. read the letter →

arxiv 2505.23501 v1 pith:24NFCALR submitted 2025-05-29 cs.RO

Masaki Murooka , Kei Okada , Masayuki Inaba This is my paper
classification cs.RO
keywords whole-bodycontactpointsearchoptimization-basedposturegenerationsmoothednormalinversekinematicshumanoidrobotsnon-prehensilemanipulationsequentialquadraticprogramming
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

This paper claims that whole-body contact postures can be generated automatically by solving for joint angles and contact points on the robot's body surface in a single gradient-based optimization. The contact point on a link is represented by two local surface coordinates, so the optimizer can move the contact point across the link's mesh while adjusting the joints. To make the search work across edges of a convex polyhedron, the paper introduces a smoothed surface normal built from weighted averages of adjacent face normals. Real-robot experiments with a dual-arm object hold, a knee push of a heavy shelf, and a two-hand board tilt show the generated postures being executed. The result matters because it removes the need to manually pre-specify every contact point for contact-rich manipulation.

What carries the argument

The load-bearing object is the local contact configuration $u=(u_1,u_2)$ with the map of eqs. (19)--(22): move on the tangent plane spanned by $\zeta,\eta$, then project onto the body's convex surface. Its Jacobian supplies the gradient of the position part of the contact task, and the smoothed normal $\hat{\xi}(p)$ of eq. (26), a normalized weighted average of the face normal and adjacent-face normals with the smooth weight of eq. (28), supplies a differentiable normal direction. Together they let a standard SQP solver search over the body surface without converting the mesh to a smooth analytic shape.

What would settle it

Run the single-body contact-point search of Fig. 7 on a box or tetrahedron from many random starting points with the target on the far side of an edge and count how often the optimization crosses the edge. If the smoothed normal does not consistently cross more often than the raw normal, or if convergence fails once the set of adjacent faces changes discretely, the central claim is falsified.

Watch

Extended reading notes

Core claim

At the center of the paper is the PN-task, a contact task whose error is zero exactly when the two contact positions coincide and their surface normals point in opposite directions. The configuration variable $\hat{q}$ stacks joint positions $\theta$ and contact configurations $u$, and the nonlinear program minimizes $\frac{1}{2}\|\hat{e}(\hat{q})\|^2$ subject to joint limits and collision constraints. Each contact configuration $u\in\mathbb{R}^2$ is local: in every SQP iteration the optimizer moves a point on the tangent plane and projects it onto the convex mesh, so the contact point itself is re-anchored each iteration rather than carried as a global coordinate. The smoothed normal of eq. (26) makes the normal and tangent directions differentiable across mesh edges, and the position Jacobian is simply the tangent axes while the normal Jacobian is evaluated by finite differences. The paper demonstrates that this formulation solves single-body contact-point searches, inverse kinematics with a shaped tip body, and whole-body motions executed on real robots.

Load-bearing premise

The load-bearing premise is that the smoothed normal of eq. (26), with derivatives approximated by finite differences, gives the SQP optimizer accurate enough gradient information to keep stepping across mesh edges instead of stalling at a face boundary.

Editorial extensions

If this is right

  • Contact points no longer need to be fixed or hand-picked; the optimizer places them automatically on any link, object, or environment surface modeled as a plane or convex polyhedron.
  • The same gradient machinery extends inverse kinematics to body-to-body contact tasks, so conventional point-to-point IK becomes a special case.
  • Because the contact configuration is re-anchored each iteration, the contact point can travel across a link during a multi-posture sequence, enabling motions such as the knee-push in which the knee contact slides on the link.
  • With the contact-wrench formulation, the method extends to statics-aware whole-body tasks, so balance and pushing forces can be considered while the knee contact point is optimized.
  • The per-iteration cost is a few milliseconds in a scripting-language implementation, so a compiled implementation is expected to be fast enough for practical motion generation.

Reading between the lines

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

  • A natural extension would be to embed this local contact-parameterization inside a trajectory optimizer, treating each contact as an implicit surface parameter that can be re-anchored as the robot moves.
  • The smoothed-normal gradient is a generic geometric primitive, so it could be reused for collision-distance gradients or for online tactile-feedback controllers beyond posture generation.
  • One testable extension is to replace the finite-difference normal Jacobian with an analytic derivative of eq. (26) and measure whether convergence across face boundaries improves or degrades when the set of adjacent faces changes discretely.
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

2 major / 5 minor

Summary. The paper proposes an optimization-based method for generating robot postures and motions that use whole-body contact. Contact points on the robot's body surface are represented by local two-dimensional configuration variables, and a sequential quadratic programming (SQP) formulation is used to simultaneously optimize joint positions and contact points. To apply gradient-based optimization to convex polyhedral meshes, the paper introduces a smoothed normal direction computed from a weighted average of adjacent face normals, with derivatives obtained by finite differences. The method is evaluated in numerical experiments on contact-point searching and inverse kinematics with surface contact, and in three real-robot whole-body contact tasks: dual-arm object holding with a PR2, knee pushing with an HRP-2, and non-prehensile board tilting with an RHP4B.

Significance. If the central claims are correct, this is a practically valuable extension of inverse kinematics to whole-body contact: it avoids explicit mesh conversion or sampling-based contact search, works directly on the raw convex mesh, and is demonstrated on physical robots in three substantially different contact tasks. The paper's strength is that the core optimization formulation is self-contained, the normal-smoothing idea is simple and plausible, and the experiments show real-world feasibility. I also credit the paper for being explicit that the global contact configuration is not carried across SQP iterations, which is an important detail for reproducibility. However, the load-bearing gradient approximation at mesh edges is only weakly validated, and the paper does not provide enough evidence that the finite-difference Jacobian is reliable when contact points cross face boundaries, which is precisely the feature that distinguishes the method from fixed-contact IK.

major comments (2)
  1. [Sec. III.B, Eq. (24)] Equation (24) states that ∂p1/∂u = [ζ η], but p^{[k+1]}(u) is defined in Eq. (22) as a projection onto the body surface. This derivative is exact only when the tangent-plane step stays within the interior of the same face; when the optimal step crosses a face boundary, the closest-point projection clamps or slides along an edge, so the true derivative differs from [ζ η] and the QP linearization in Eq. (11) may over- or under-estimate the effect of the contact configuration update. The paper does not bound this error, nor does it report how often edge-crossing steps occur or compare the finite-difference Jacobian against a high-resolution or analytic reference. Since the method's claimed advantage is precisely the ability to move contact points across mesh edges, this approximation needs either a derivation of a correction/error bound or a systematic numerical study (e.g., repeated trials with varied initial conditions, counts of edge-crossing steps, and convergence rates) to be considered load-bearing evidence.
  2. [Sec. III.C, Eq. (25)-(28)] The smoothed normal of Eq. (26) is defined by a weighted sum over the set of adjacent faces Ψ(ψ), and the text acknowledges in Sec. III.C that this face set changes discretely. Nevertheless, the paper claims the smoothed normal is differentiable and computes its derivative by finite differences in Eq. (25) without analyzing the effect of the discrete face-set change. The finite-difference step ε and the smoothing range R are free parameters, and their influence on convergence is not investigated. The evidence provided—Fig. 7(C) showing failure without smoothing and Figs. 7(A)-(B), 11-15 showing success with smoothing—is limited to a small number of qualitative trajectories; no convergence statistics, sensitivity analyses, or failure counts are reported. I request either a continuity/differentiability analysis that covers the face-set boundary or a more extensive experimental study (varying R, ε, ulim, and initial contact points) to support the claim that the finite-difference Jacobian is adequate across edges.
minor comments (5)
  1. [Sec. III.C, Eq. (27)] The distance x(p, ψ̄) is described only as "the distance on the body surface from point p to the adjacent face"; please specify whether this is Euclidean distance to the face, shortest path along the surface, or another geodesic-like measure, since the formula and its implementation affect reproducibility.
  2. [Sec. IV.A] The text states that with a target not on the body surface, the closest contact point is found, but no error metric or quantitative result is reported for that case; please add a numerical value or a convergence plot.
  3. [Table I] The percentage breakdown in Table I is confusing: "calculate matrices in QP (11)" is listed as 82% and 87% of total time, while its subrows (∂e/∂θ, ∂e/∂u, c(q), solve QP) sum to only about 87-99% of the total; please clarify the hierarchy of the timing categories so the percentages sum consistently.
  4. [Sec. V.A] The assertion that a 3 kg object is difficult to grasp with just the gripper tips due to insufficient friction is not supported by a quantitative friction or force-closure analysis; a brief justification or citation would help.
  5. [General] There are minor typographical issues, including "Mechano-Infomatics" in the author affiliation and inconsistent spacing in equations; a careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

Central contact-point-search derivation is self-contained; self-citations are not load-bearing.

full rationale

The paper's central claim is that a gradient-based optimization can simultaneously search joint positions and contact points parameterized by local two-dimensional coordinates on a convex mesh. The derivation chain is self-contained: Eqs. (19)-(22) define the local tangent-plane motion and closest-point projection, Eqs. (23)-(25) build the PN-task Jacobian from that same map, and Eq. (26) defines the smoothed normal used for differentiability. The numerical validation in Sec. IV-A compares the optimizer against externally specified target points and normals, not against quantities derived from the method's own parameters, so no fitted input is being renamed as a prediction. The weights w_nrm and smoothing range R are user-chosen constants, not data-fit parameters. The only self-citations, [1] and [21], appear in the related-work discussion and in the extension to an inverse-kinematics-and-statics framework for the knee-pushing application; neither supplies or justifies the paper's distinguishing contribution (contact point search on the body surface by gradient-based optimization), and neither is invoked as a uniqueness theorem or as an external constraint that forces the formulation. The approximation in Eq. (24) and the finite-difference Jacobian in Eq. (25) are numerical accuracy concerns, not circularities, because they are not equivalent to the paper's conclusions by construction. Overall, no significant circularity is present.

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

The method's central claim depends only on user-chosen smoothing and weight parameters plus a local convex-geometry parameterization; no fitted constants or new physical entities are introduced. The main heuristic is the differentiability assumption for the smoothed normal.

free parameters (3)
  • smoothing range R = 400 mm (Fig. 4), 30 mm (Fig. 6)
    User-selected range over which adjacent face normals are blended in eq. (26); affects the shape of the smoothed normal and hence the optimization landscape, but is not fitted to any data.
  • normal weight wnrm = 1 and 10 (Fig. 7 B1, B2)
    Relative weight between positional and normal components of the PN-task in eq. (7); chosen per experiment, not fitted.
  • contact configuration limit ulim = 10-20 mm for life-sized robots
    Limits the tangent-plane step in eq. (21) to keep the local linearization valid; set by the user based on robot scale.
assumptions (3)
  • domain assumption Robot links, objects, and environments are modeled as convex polyhedra or primitive shapes.
    Problem definition in Sec. II.A limits bodies to planes and convex polyhedra; real robot links are used via their convex hull, which may discard concave features.
  • ad hoc to paper The smoothed normal in eq. (26) is differentiable enough for gradient-based optimization, with derivatives computed by finite differences in eq. (25).
    The paper introduces this heuristic to handle edge non-differentiability of raw polyhedra; the finite-difference approximation is asserted, not proved, to give usable gradients across face boundaries.
  • standard math The closest-point projection onto a convex body is well-defined and locally well-behaved; the tangent-plane linearization of eqs. (19)-(22) is a valid local parameterization within the ulim bound.
    Convexity ensures a unique closest point; the paper uses this mapping for contact point updates (Sec. III.A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimization-based Posture Generation for Whole-body Contact Motion by Contact Point Search on the Body Surface." pith.science (2026). https://pith.science/paper/24NFCALR

@misc{pith2026250523501,
  author       = {Pith},
  title        = {Pith review of: Optimization-based Posture Generation for Whole-body Contact Motion by Contact Point Search on the Body Surface},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24NFCALR}},
  note         = {Machine review of arXiv:2505.23501}
}
read the original abstract

Whole-body contact is an effective strategy for improving the stability and efficiency of the motion of robots. For robots to automatically perform such motions, we propose a posture generation method that employs all available surfaces of the robot links. By representing the contact point on the body surface by two-dimensional configuration variables, the joint positions and contact points are simultaneously determined through a gradient-based optimization. By generating motions with the proposed method, we present experiments in which robots manipulate objects effectively utilizing whole-body contact.

Figures

Figures reproduced from arXiv: 2505.23501 by the authors.

Figure 2
Figure 2. Example of posture generation. (A) The figure shows an example of the whole-body contact posture generated by the proposed method. The robot is in contact with the environment in three places. The contacts of knee and elbow are des￾ignated as “link-to-plane” target, whereas the foot contact is designated as a conventional “point-to-point” target. (B) The figure also displays the enlarged view of contact at the knee.… view at source ↗
Figure 1
Figure 1. Examples of body to be contact. The body is modeled as a plane or convex polyhedron. For robot links, we use the convex polyhedron generated from the detailed mesh model automatically. Although the number of vertices in the body is high (about 50), the proposed optimization method based on normal smoothing works properly. We use simple primitive shapes (cylinder, cube, and sphere.) for the object and environment. (A… view at source ↗
Figure 3
Figure 3. Mapping from the contact configuration to the contact point. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Example of a normal calculation of the body surface. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Smooth interpolation function (R = 1). IV. NUMERICAL EXAMPLES We implemented the proposed motion generation algo￾rithm with EusLisp [17], that is an extension of the Common Lisp for robot programming. We used external libraries qpOASES [18] and Bullet [19] for QP and c…
Figure 6
Figure 6. Figure 6: Example of normal calculation of the 3D convex mesh. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Inverse kinematics with surface contact on tip body. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: Comparison of solvability with and without a tip body. [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 13
Figure 13. Figure 13: Since a two-dimensional configuration representing [PITH_FULL_IMAGE:figures/full_fig_p006_13.png]
Figure 10
Figure 10. Figure 10: Error in position and normal direction. The baseline method (project) and the proposed method (gradient) are compared. The normal direction error of the proposed method turned to zero after 15 iterations because of computer calculation accuracy. high owing to the many…
Figure 12
Figure 12. Figure 12: Contact points in dual-arm holding posture. [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]
Figure 13
Figure 13. Figure 13: History of contact points in posture generation. [PITH_FULL_IMAGE:figures/full_fig_p007_13.png]
Figure 15
Figure 15. Figure 15: Tilting operation performed on a large board using both hands. [PITH_FULL_IMAGE:figures/full_fig_p008_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    Murooka, S

    M. Murooka, S. Nozawa, Y. Kakiuchi, K. Okada, and M. Inaba. Whole- body pushing manipulation with contact posture planning of large and heavy object for humanoid robot. InProceedings of The 2015 IEEE International Conference on Robotics and Automation, pp. 5682–5689, May 2015

    work page 2015

  2. [2]

    Kanehiro, K

    F . Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, S. Kajita, K. Yokoi, H. Hirukawa, K. Akachi, and T . Isozumi. The first humanoid robot that has the same size as a human and that can lie down and get up. InProceedings of The 2003 IEEE International Conference on Robotics and Automation, Vol. 2, pp. 1633–1639, Sep. 2003

    work page 2003

  3. [3]

    Mittendorfer, E

    P . Mittendorfer, E. Yoshida, and G. Cheng. Realizing whole-body tactile interactions with a self-organizing, multi-modal artificial skin on a humanoid robot.Advanced Robotics, Vol. 29, No. 1, pp. 51–67, 2015

    work page 2015

  4. [4]

    Kakiuchi, M

    Y. Kakiuchi, M. Kamon, N. Shimomura, S. Yukizaki, N. Takasugi, S. Nozawa, K. Okada, and M. Inaba. Development of life-sized humanoid robot platform with robustness for falling down, long time working and error occurrence. InProceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 689– 696, Sept 2017

    work page 2017

  5. [5]

    Bouyarmane, S

    K. Bouyarmane, S. Caron, A. Escande, and A. Kheddar.Humanoid Robotics: A Reference, chapter Multi-contact Motion Planning and Control, pp. 1763–1804. Springer, January 2018

    work page 2018

  6. [6]

    Werner, B

    A. Werner, B. Henze, D. A. Rodriguez, J. Gabaret, O. Porges, and M. A. Roa. Multi-contact planning and control for a torque-controlled humanoid robot. InProceedings of the 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5708–5715, Oct 2016

    work page 2016

  7. [7]

    Salisbury, W

    K. Salisbury, W . Townsend, B. Ebrman, and D. DiPietro. Preliminary design of a whole-arm manipulation system (wams). InProceedings of The 1988 IEEE International Conference on Robotics and Automation, Vol. 1, pp. 254–260, April 1988

    work page 1988

  8. [8]

    Miller and P

    A. Miller and P . K. Allen. Graspit!: A Versatile Simulator for Robotic Grasping.IEEE Robotics and Automation Magazine, Vol. 11, No. 4, pp. 110–122, 2004

    work page 2004

Show all 24 references
  1. [9]

    M. A. Roa, M. J. Argus, D. Leidner, C. Borst, and G. Hirzinger. Power grasp planning for anthropomorphic robot hands. InProceedings of The 2012 IEEE International Conference on Robotics and Automation, pp. 563–569, May 2012

    2012

  2. [10]

    H. Dai, A. Valenzuela, and R. Tedrake. Whole-body motion planning with centroidal dynamics and full kinematics. InProceedings of the 2014 IEEE-RAS International Conference on Humanoid Robots, pp. 295–302, Nov 2014

    2014

  3. [11]

    Escande, N

    A. Escande, N. Mansard, and P . B. Wieber. Hierarchical quadratic programming: Fast online humanoid-robot motion generation.The International Journal of Robotics Research, Vol. 33, No. 7, pp. 1006– 1028, 2014

    2014

  4. [12]

    Brossette, A

    S. Brossette, A. Escande, and A. Kheddar. Multicontact postures computation on manifolds.IEEE Transactions on Robotics, Vol. 34, No. 5, pp. 1252–1265, Oct 2018

    2018

  5. [13]

    Escande, S

    A. Escande, S. Miossec, M. Benallegue, and A. Kheddar. A strictly convex hull for computing proximity distances with continuous gradients.IEEE Transactions on Robotics, Vol. 30, No. 3, pp. 666–678, June 2014

    2014

  6. [14]

    Escande, S

    A. Escande, S. Brossette, and A. Kheddar. Parametrization of catmull- clark subdivision surfaces for posture generation. InProceedings of The 2016 IEEE International Conference on Robotics and Automation, pp. 1608–1614, May 2016

    2016

  7. [15]

    D. E. Whitney. Resolved motion rate control of manipulators and human prostheses.IEEE Transactions on man-machine systems, Vol. 10, No. 2, pp. 47–53, 1969

    1969

  8. [16]

    Nocedal and S

    J. Nocedal and S. Wright. Numerical optimization.Springer, 1999

    1999

  9. [17]

    EusLisp.http://github.com/euslisp

  10. [18]

    H. J. Ferreau, C. Kirches, A. Potschka, H. G. Bock, and M. Diehl. qpOASES: A parametric active-set algorithm for quadratic program- ming.Mathematical Programming Computation, Vol. 6, No. 4, pp. 327–363, 2014

    2014

  11. [19]

    Coumanns

    E. Coumanns. Bullet physics library.www.bulletphysics.org, 2012

    2012

  12. [20]

    Bouyarmane and A

    K. Bouyarmane and A. Kheddar. Humanoid robot locomotion and manipulation step planning.Advanced Robotics, Vol. 26, No. 10, pp. 1099–1126, 2012

    2012

  13. [21]

    Shigematsu, M

    R. Shigematsu, M. Murooka, Y. Kakiuchi, K. Okada, and M. Inaba. Generating a key pose sequence based on kinematics and statics optimization for manipulating a heavy object by a humanoid robot. InProceedings of the 2019 IEEE/RSJ International Conference on Intelligent Robots an...

    2019

  14. [22]

    Kajita, M

    S. Kajita, M. Morisawa, K. Miura, S. Nakaoka, K. Harada, K. Kaneko, F . Kanehiro, and K. Yokoi. Biped walking stabilization based on linear inverted pendulum tracking. InProceedings of the 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 4489– 4496...

    2010

  15. [23]

    Nozawa, et al

    S. Nozawa, et al. Multi-layered real-time controllers for humanoid’ s manipulation and locomotion tasks with emergency stop. InPro- ceedings of the 2015 IEEE-RAS International Conference on Humanoid Robots, pp. 381–388, Nov 2015

    2015

  16. [24]

    M. T . Mason. Progress in nonprehensile manipulation.The Interna- tional Journal of Robotics Research, Vol. 18, No. 11, pp. 1129–1141, 1999

    1999

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.