Proactive tactile exploration for object-agnostic shape reconstruction from minimal visual priors

Costas S. Tzafestas; George Retsinas; Paris Oikonomou; Petros Maragos

REVIEW 3 major objections 6 minor 34 references

Proactive tactile exploration for object-agnostic shape reconstruction from minimal visual priors

T0 review · 3 major / 6 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 A robot can reconstruct an unknown object's whole surface from a small single-view point cloud and a few cautious touches, by treating the object as a deformed ellipsoid and picking each touch to shrink uncertainty without missing the…

desk verdict A genuinely new contact-failure-aware exploration objective buried under an over-broad 'object-agnostic' claim and cherry-picked baseline reporting. read the letter →

arxiv 2505.11975 v1 pith:VJQUTWXY submitted 2025-05-17 cs.RO

Paris Oikonomou , George Retsinas , Petros Maragos , Costas S. Tzafestas This is my paper · ORCID

classification cs.RO

keywords 3Dshapereconstructionactivetactileexplorationvisuo-hapticperceptionellipsoidtemplatethin-plate-splinedeformationuncertaintypropagationcontactfailurerobotmanipulation

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 aims to show that a robot that sees only a small patch of an unknown object—say, a partial view from one camera—can still reconstruct the entire 3D surface by combining that visual hint with a small number of tactile pokes. The proposed pipeline first fits a generic ellipsoid template to the sparse visual points to get a rough global shape, then deforms the mesh locally with thin-plate-spline interpolation to match the incoming tactile points. Around this, a proactive exploration rule chooses the next touch point by balancing two goals: moving toward unexplored, high-uncertainty regions while staying close to already-confident surface area, so the robot rarely pokes into empty space. In simulation, after 50 contacts the reconstructed meshes reach chamfer distances below 4 mm for all tested objects, and on a real arm the procedure works with a depth camera and force-torque tactile sensors. If this holds, robots could inspect or grasp objects they have barely seen, using only a cheap camera and a few deliberate contacts.

What carries the argument

The load-bearing component is the two-step coarse-to-fine mesh fitting together with the exploration objective in Eq. (3). Global shape is estimated by gradient-descent fitting of an ellipsoid template, with parameters rotation $R$, translation $t$ and scales $s$, to the attractor points through a least-squares loss. Local shape is then resolved by a thin-plate-spline interpolator $F$ that displaces each mesh vertex along its normal, so the mesh can represent bumps and dents. Uncertainty from each attractor is spread to nearby vertices with a heavy-tailed weight $1/(1+t^2)$, and the next contact is the vertex $w$ maximizing $\alpha_G G_j + \alpha_U U_j$ subject to having at least one 'confident' neighbor and uncertainty above a threshold. The force/torque-based uncertainty of Eq. (1) turns each tactile reading into a reliability score, and the whole loop iterates until the uncertainty map is sufficiently low.

What would settle it

Run the same pipeline on a torus or a box with a handle, using the same 50-contact budget and evaluation metrics; if the chamfer distance fails to drop below 4 mm or the contact-failure rate stays high, the claim that a single ellipsoid prior suffices for object-agnostic reconstruction is falsified. A simpler observational test is to watch whether the reconstructed mesh ever creates a hole: the thin-plate-spline deformation cannot change topology, so a successful torus reconstruction would require a mechanism the paper does not provide.

Watch

Extended reading notes

Core claim

The central claim is that a single ellipsoid template, deformed in two stages, plus a cautious active-touch policy is enough to recover the shape of common objects from almost no visual information. The key discovery is that the exploration strategy—not the fitting subroutine—does the heavy lifting: choosing touch points that maximize a weighted sum of geodesic distance from confident vertices and local uncertainty, subject to having at least one confident neighbor, drives a depth-first traversal of the surface that discovers the global structure while keeping contact failures low. The paper reports that for the most challenging objects in its test set (a coffee can, a sugar box, a mustard container), this strategy achieves reconstruction errors substantially below a standard uncertainty-only baseline over the same 50 contact attempts—for example, 3.82 mm versus 9.97 mm chamfer distance for the sugar box—with far fewer failed pokes. The same pipeline also works on a real robot, reconstructing a 3D-printed mustard container to 4.1 mm chamfer error in 30 iterations despite the noise of a real depth camera and tactile sensor.

Load-bearing premise

The whole approach stands on the assumption that the target object's shape is roughly ellipsoidal—globally convex with no through-holes—so that a single ellipsoid fit gets close enough to the real surface for the local deformation stage to finish the job.

Editorial extensions

If this is right

After 50 contacts, all simulated objects—including a coffee can, a sugar box, and a mustard container—reach chamfer distances below 4 mm, with no prior object model beyond the ellipsoid assumption.
The proactive strategy cuts contact failures by roughly a factor of two to three compared with uncertainty-only exploration on the hardest objects, and it avoids the failure loops that can stall the baseline.
The method transfers to a real 7-DoF arm with a depth camera and force/torque tactile sensors, reconstructing a 3D-printed mustard container to 4.1 mm chamfer error in 30 iterations.
The output includes a per-vertex uncertainty map, which tells a downstream manipulation system which regions of the surface are still unexplored and therefore where additional touch or visual data would be most valuable.

Reading between the lines

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

A comparable cautious depth-first policy could be applied to other active-perception problems, such as tracing object boundaries or segmenting a cluttered scene, but the present paper only demonstrates it for single-object surface reconstruction.
The fixed ellipsoid template is the likely ceiling: an object with a through-hole or sharp concavity would break the global fit, so a natural next step is online selection among multiple templates, which the paper only lists as future work.
The paper's contact-failure metric (a 15 mm deviation between estimated and real contact) is a geometric proxy; in real manipulation, failures also depend on approach speed, friction, and surface compliance, so the risk reduction may not transfer unchanged to dynamic tasks.
Because visual points are assigned a fixed high uncertainty, the method is conservative near the camera view; a testable tweak would be to set visual uncertainty per-pixel from the depth sensor's noise model, which could speed up exploration on visually reliable regions.

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.

Desk Editor's Note

read the letter

The thing to know: this paper has a real idea. The exploration objective in Eq. 3, which balances geodesic distance from confident regions against uncertainty while requiring a confident neighbor, is not just a tweak on min-uncertainty. It demonstrably cuts contact failures on the tested objects, and the two-step ellipsoid-plus-TPS pipeline is a sensible way to turn a handful of points into a usable mesh. The tactile uncertainty model in Eq. 1 is sensor-grounded, not circular. Credit where due: the comparison against minU is on the right axis, the chamfer numbers are plausible, and the real-robot demo on the mustard container shows the loop actually closes outside simulation.

Now the soft spots, in order of size. First, the title and abstract say 'object-agnostic,' but the method assumes a deformed ellipsoid. The local TPS displacement moves vertices along their original normals; it cannot create holes or handles. The evaluation sticks to genus-0, roughly ellipsoidal shapes, with the sugar box as the most extreme case. The paper's own Future Work admits a template selection strategy would help. That's a load-bearing overclaim. The method is a good fit for mildly deformed ellipsoids; calling it object-agnostic is not supported.

Second, Table I reports only the minU runs that 'made relatively significant progress,' as the text admits. That is selection bias, and it flatters the comparison. The authors should report all runs, including the infinite-loop failures, or at least give a worst-case bound. Third, the exploration weights alpha_G, alpha_U and the uncertainty thresholds u'_min, u'_max are never given. Without those values the reader cannot reproduce Eq. 3, and sensitivity to them is unknown. That is a minor fix but a necessary one. Fourth, the real-world validation is a single object, which is fine for a preliminary demo but not for the generalization claim.

On the citation pattern: the related work is appropriately broad, and I did not see self-citation padding. The circularity concern raised by the skeptic does not land; the uncertainty comes from the sensor measurement, not from the reconstruction output.

Bottom line: the core contribution — a cautious, failure-aware exploration strategy for visuo-tactile shape reconstruction — is solid and worth publishing after serious revisions. Whoever referees this should push hard on the object-agnostic language and the baseline reporting, but they should not dismiss the method. I would bring it to a reading group that works on active perception, and I would likely cite it once the hyperparameters are released. Send it to peer review, not to desk reject.

Referee Report

3 major / 6 minor

Summary. The paper proposes an iterative visuo-tactile pipeline for 3D object shape reconstruction from a minimal single-view point cloud. Starting from this visual prior, the method fits a single ellipsoid template to the sparse attractors, refines the surface locally with a thin-plate-spline deformation, propagates sensor-derived uncertainty to mesh vertices, and then selects the next tactile contact by maximizing a weighted combination of geodesic distance from confident vertices and uncertainty, subject to confidence constraints (Eq. 3). The approach is evaluated in simulation on seven objects (lamp, wine bottle, and five YCB objects) and on a real robotic setup with a 3D-printed mustard container, using chamfer distance and number of contact failures as metrics and comparing against a minU baseline that minimizes uncertainty without failure awareness.

Significance. If the results hold, the paper offers a practical contribution: it shows that a reasonably accurate surface can be recovered with a small number of tactile contacts and a deliberately failure-avoiding exploration policy, and it validates the idea on a real robot. The explicit uncertainty model for tactile measurements (Eq. 1) and the real-robot demonstration are strengths, as is the inclusion of a quantitative baseline comparison. However, the central 'object-agnostic' claim is broader than the evidence: all tested objects are genus-0 and approximately ellipsoidal, and the single-template deformation cannot represent holes, handles, or deep undercuts. The comparison with minU also excludes some failed baseline runs, which inflates the apparent advantage. With a scoped claim and a more complete comparison, the method would be a solid contribution to active visuo-tactile perception.

major comments (3)

[III-A, III-D, III-E; title and abstract] The central claim of object-agnostic reconstruction is not supported by the single-ellipsoid template. The local deformation in Section III-E moves each vertex along its original normal (Eq. 2), so the mesh cannot change genus or create holes/handles; Section III-A states that only an ellipsoid template is considered. The evaluation in Table I and Figs. 4-5 covers the lamp, wine bottle, pear, ball, coffee can, mustard container, and sugar box, all genus-0 objects that are qualitatively ellipsoidal. The paper's own Future Work (Section V) concedes that template selection would be needed for broader applicability. Please either replace the term 'object-agnostic' with a scope-limited description (e.g., 'ellipsoid-like objects') throughout, or add experiments with non-ellipsoidal topologies (objects with holes, handles, or deep undercuts) and discuss failure modes.
[Table I and Section IV-B] The quantitative comparison to minU is affected by selection bias. The text states that 'the results presented in Table I concern only the cases where minU strategy managed to make relatively significant progress in reconstruction,' and that minU was sometimes trapped in an infinite loop of contact failures. Excluding those runs removes the worst outcomes of the baseline, yet Table I still reports mean and standard deviation over 'three executions' without indicating how many runs were discarded. Please report all runs, including the number of discarded/trapped cases, or use a metric that penalizes divergence, so that the comparison is conservative and reproducible.
[Section III-G, Eq. (3), and Section IV] The exploration strategy and its evaluation depend on several parameters that are not reported: the weights alpha_G and alpha_U, the thresholds u'_min and u'_max, the maximum tactile uncertainty u_max, the visual-attractor uncertainty (given as 0.4 only for the real experiment), and the traverse threshold for uncertainty propagation. Without these values (or a sensitivity study), the simulation results cannot be reproduced and it is unclear how robust the improvement over minU is to parameter choice. Please include a parameter table and, ideally, an analysis of how the chamfer distance and failure count vary with the weights and thresholds.

minor comments (6)

[Fig. 6 and Table I] The caption of Fig. 6 says the plots correspond to the mean of chamfer distances after five independent executions, while the caption of Table I says the mean and standard deviation are after three executions; please reconcile these numbers.
[IV-A and IV-C] The size of the visual prior is not quantified (number of points, angular coverage, or percentage of the object surface). Since 'minimal visual priors' is a central claim, please report this quantity for the simulation and real experiments.
[III-G] The definition of the geodesic distance G_j and the 'confident vertices of its neighborhood' used in Eq. (3) is informal. Please specify the graph construction (mesh edges, k-ring, or other), how the mean geodesic distance is computed, and how the confidence thresholds u'_min and u'_max are chosen.
[III-C] In Eq. (1), please state what happens when F_z,i is near zero, and whether the force and torque measurements are filtered or averaged over the contact duration before computing the uncertainty.
[IV-B] The phrase 'for the same number of contact attempts' should clarify that contact-failure iterations consume an attempt but do not update the reconstruction, so OURS and minU receive different numbers of successful contacts even at the same iteration count.
[Fig. 4] The caption says 'the first two objects are examined here,' but the third row shows the real mustard container from Section IV-C; please clarify that the third row corresponds to the real-robot experiment.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: external ground-truth evaluation; ellipsoid assumption is a scope limitation, not a self-referential derivation.

full rationale

The derivation chain is not circular. The reconstruction pipeline fits an ellipsoid template to visual and tactile attractors (Section III-D) and then applies a thin-plate-spline interpolation so that each attractor satisfies Eq. 2, with vertex displacements computed from the fitted interpolator (Section III-E). Attractor uncertainties come from sensor measurements via Eq. 1 and are propagated to mesh vertices by proximity weighting (Section III-F); none of these equations defines the reported chamfer distance. The quantitative evaluation in Table I compares the reconstructed meshes against external ground-truth object meshes after 50 iterations, and the comparison against the minU baseline uses identical reconstruction parameters and initial conditions, differing only in the exploration objective (Eq. 3). That objective is a weighted combination of propagated uncertainty and geodesic distance; it is a heuristic for choosing contacts, not the evaluation metric, so the lower chamfer error reported for the proposed strategy is an externally measured outcome rather than a consequence of definition. The reference list contains no self-citations by the authors, and no uniqueness theorem is imported. The main caveat is the ellipsoid-template assumption stated in Section II ('a wide variety of shapes and objects could be seen as mildly or severely deformed ellipsoids') and operationalized in Section III-D: since the local deformation step displaces vertices along their normals, it cannot introduce holes, handles, or other non-ellipsoidal topology, so the 'object-agnostic' title overstates the tested scope. The authors themselves acknowledge this in Section V, noting that 'a template selection strategy could choose among multiple templates in real-time.' That is a correctness and generalization limitation, not circularity: the claimed reconstruction results are not encoded in the inputs by construction. Therefore no circular step is identified, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method leans on several domain assumptions rather than invented physical entities. The most load-bearing is the deformed-ellipsoid shape prior, which bounds the class of objects the method can handle. The uncertainty model and exploration objective are heuristic and enter through hand-tuned parameters. The paper does not introduce new physical entities, and the term 'attractors' is a label for measured surface points, not a new ontological category.

free parameters (6)

exploration weights alpha_G, alpha_U = not reported
Weights in Eq. 3 that balance geodesic distance and uncertainty in the contact-point selection. The paper does not state their values or sensitivity analysis.
uncertainty thresholds u'_min, u'_max = not reported
Constraints in Eq. 3 that require the candidate vertex to be above a minimum uncertainty and adjacent to a confident vertex. Values are chosen by hand and not disclosed.
maximum tactile uncertainty u_max = not reported
Scaling factor in Eq. 1 that determines the upper bound of tactile attractor uncertainty. Value is not specified.
visual attractor uncertainty = 0.4 (real experiment)
Fixed uncertainty assigned to all visual points in the real-robot experiment (Section IV-C). In simulation, the value is not stated.
contact failure threshold = 15 mm
Simulation criterion for contact failure defined in Section IV-B: distance between candidate point on estimated surface and intersection with real surface exceeds 15 mm.
traverse threshold = 5 edges
Propagation neighborhood size used when updating vertex uncertainties from an attractor (Section III-F). Chosen as an example, not justified.

assumptions (5)

domain assumption Target objects can be adequately approximated as deformed ellipsoids.
Stated in Section III-A and used in the global shape estimation step (Section III-D) where an ellipsoid template is fitted to visual points. If the object has a topology far from an ellipsoid, the global fit may be poor.
domain assumption The single-view depth camera provides a small but sufficient point cloud to locate the object and initialize the ellipsoid fit.
Assumed throughout the scenario description (Sections I and III-A). The method does not address cases where the visual prior is too noisy or too small to give a reasonable initialization.
domain assumption The uncertainty propagation heuristic (multiplying attractor uncertainty by 1/(1+t^2) and keeping the minimum for each vertex) accurately reflects the spatial uncertainty of the reconstructed surface.
Section III-F defines this propagation without empirical or theoretical justification. The exploration strategy relies on these uncertainty scores to select the next contact point.
domain assumption The exploration objective in Eq. 3, maximizing weighted geodesic distance from confident vertices plus uncertainty, is an effective proxy for minimizing total reconstruction uncertainty with few contacts.
Section III-G introduces this heuristic. The paper provides no proof or ablation that this objective is optimal or near-optimal compared to other exploration criteria.
domain assumption The tactile sensor uncertainty formula (Eq. 1), derived from measured torques and forces, correctly models the reliability of each contact point.
Section III-C asserts that torque around x and y axes and force on the z axis determine the contact location uncertainty. This relationship is stated without experimental calibration or validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Proactive tactile exploration for object-agnostic shape reconstruction from minimal visual priors." pith.science (2026). https://pith.science/paper/VJQUTWXY

@misc{pith2026250511975,
  author       = {Pith},
  title        = {Pith review of: Proactive tactile exploration for object-agnostic shape reconstruction from minimal visual priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJQUTWXY}},
  note         = {Machine review of arXiv:2505.11975}
}

read the original abstract

The perception of an object's surface is important for robotic applications enabling robust object manipulation. The level of accuracy in such a representation affects the outcome of the action planning, especially during tasks that require physical contact, e.g. grasping. In this paper, we propose a novel iterative method for 3D shape reconstruction consisting of two steps. At first, a mesh is fitted on data points acquired from the object's surface, based on a single primitive template. Subsequently, the mesh is properly adjusted to adequately represent local deformities. Moreover, a novel proactive tactile exploration strategy aims at minimizing the total uncertainty with the least number of contacts, while reducing the risk of contact failure in case the estimated surface differs significantly from the real one. The performance of the methodology is evaluated both in 3D simulation and on a real setup.

Figures

Figures reproduced from arXiv: 2505.11975 by the authors.

**Figure 1.** Overview of the proposed pipeline. To ease the comprehension of the methodology, the shape reconstruction of a 2d bottleshape object is considered, where a path graph is used, instead of a 3d mesh. The colored blocks correspond to each of the individual algorithmic components taking part in the reconstruction (details in Section III). In each block, the top instance illustrates the outcome of the process during the… view at source ↗

**Figure 2.** Top left: The tactile sensor is a variant of the Shokac Chip 6DoF-P18 by Touchence [31] and measures the force/torque applied along all three axes. Bottom left: At each iteration, the robot moves within a predefined distance d before (red) and after (blue) the candidate contact point (green) along its estimated normal vector. In case of no contact, the attempt is considered as failure. Rest with “X”: Other cases of … view at source ↗

**Figure 3.** First three iterations of the proposed active tactile exploration strategy (Section III-G). At the side of the visual points (cyan), the reconstructed mesh (red-green) fits the actual one (grey). At the other side tactile exploration is required. Following the proposed algorithm, a path of tactile points (blue) is created from a neighborhood close to the visual data up to the unexplored side of the surface, resembl… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Iterative progression of the proposed methodology. The first two rows corresponds to the evaluation performed in simulation, while the last one took place on the real experimental setup (Sec. IV-C). From left to right: target mesh, 6 instances of the reconstruction at …

**Figure 5.** Figure 5: Reconstruction of several objects (from left to right: coffee can, box of sugar, mini soccer ball, pear, mustard container) from YCB dataset, after 50 iterations of the methodology (including contact failure iterations). Top: Real mesh of the object and collected attr…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

[1]

Humans integrate visual and haptic information in a statistically optimal fashion,

M. O. Ernst and M. S. Banks, “Humans integrate visual and haptic information in a statistically optimal fashion,”Nature, vol. 415, no. 6870, pp. 429–433, 2002

2002
[2]

Deepsdf: Learning continuous signed distance functions for shape representation,

J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “Deepsdf: Learning continuous signed distance functions for shape representation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

work page 2019
[3]

Shape completion enabled robotic grasping,

J. Varley, C. DeChant, A. Richardson, J. Ruales, and P. Allen, “Shape completion enabled robotic grasping,” in2017 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2017, pp. 2442–2447

work page 2017
[4]

Robust grasp planning over uncertain shape completions,

J. Lundell, F. Verdoja, and V . Kyrki, “Robust grasp planning over uncertain shape completions,” in2019 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 1526–1532

work page 2019
[5]

Completion and reconstruction with primitive shapes,

R. Schnabel, P. Degener, and R. Klein, “Completion and reconstruction with primitive shapes,”Computer Graphics F orum, vol. 28, no. 2, pp. 503–512, 2009. [Online]. Available: https://onlinelibrary.wiley.com/ doi/abs/10.1111/j.1467-8659.2009.01389.x

arXiv 2009
[6]

Active haptic perception in robots: a review,

L. Seminara, P. Gastaldo, S. J. Watt, K. F. Valyear, F. Zuher, and F. Mastrogiovanni, “Active haptic perception in robots: a review,” Frontiers in neurorobotics, vol. 13, p. 53, 2019

work page 2019
[7]

Active tactile exploration with uncer- tainty and travel cost for fast shape estimation of unknown objects,

T. Matsubara and K. Shibata, “Active tactile exploration with uncer- tainty and travel cost for fast shape estimation of unknown objects,” Robotics and Autonomous Systems, vol. 91, pp. 314–326, 2017

work page 2017
[8]

Active perception: Building objects’ models using tactile exploration,

N. Jamali, C. Ciliberto, L. Rosasco, and L. Natale, “Active perception: Building objects’ models using tactile exploration,” in2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids), 2016, pp. 179–185

work page 2016

Show all 34 references

[9]

Active tactile object exploration with gaussian processes,

Z. Yi, R. Calandra, F. Veiga, H. van Hoof, T. Hermans, Y . Zhang, and J. Peters, “Active tactile object exploration with gaussian processes,” in2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2016, pp. 4925–4930

2016
[10]

Active learning with query paths for tactile object shape exploration,

D. Driess, P. Englert, and M. Toussaint, “Active learning with query paths for tactile object shape exploration,” in2017 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), 2017, pp. 65–72

2017
[11]

Iterative closest labeled point for tactile object shape recognition,

S. Luo, W. Mou, K. Althoefer, and H. Liu, “Iterative closest labeled point for tactile object shape recognition,” in2016 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), 2016, pp. 3137–3142

2016
[12]

Ergodic exploration using binary sensing for nonparametric shape estimation,

I. Abraham, A. Prabhakar, M. J. Hartmann, and T. D. Murphey, “Ergodic exploration using binary sensing for nonparametric shape estimation,”IEEE robotics and automation letters, vol. 2, no. 2, pp. 827–834, 2017

2017
[13]

Multi-contact haptic exploration and grasping with tactile sensors,

N. Sommer and A. Billard, “Multi-contact haptic exploration and grasping with tactile sensors,”Robotics and autonomous systems, vol. 85, pp. 48–61, 2016

2016
[14]

Online active and dynamic object shape exploration with a multi-fingered robotic hand,

F. Khadivar, K. Yao, X. Gao, and A. Billard, “Online active and dynamic object shape exploration with a multi-fingered robotic hand,” Robotics and Autonomous Systems, vol. 166, p. 104461, 2023

2023
[15]

Visuo- haptic object perception for robots: an overview,

N. Navarro-Guerrero, S. Toprak, J. Josifovski, and L. Jamone, “Visuo- haptic object perception for robots: an overview,”Autonomous Robots, vol. 47, no. 4, pp. 377–403, 2023

2023
[16]

Enhancing visual perception of shape through tactile glances,

M. Bj ¨orkman, Y . Bekiroglu, V . H¨ogman, and D. Kragic, “Enhancing visual perception of shape through tactile glances,” in2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2013, pp. 3180–3186

2013
[17]

3d shape perception from monocular vision, touch, and shape priors,

S. Wang, J. Wu, X. Sun, W. Yuan, W. T. Freeman, J. B. Tenenbaum, and E. H. Adelson, “3d shape perception from monocular vision, touch, and shape priors,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1606– 1613

2018
[18]

Implicit geometric regularization for learning shapes,

A. Gropp, L. Yariv, N. Haim, M. Atzmon, and Y . Lipman, “Implicit geometric regularization for learning shapes,” inProceedings of the 37th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, H. D. III and A. Singh, Eds., vol
[19]

Visual- tactile fusion for 3d objects reconstruction from a single depth view and a single gripper touch for robotics tasks,

M. Tahoun, O. Tahri, J. A. Corrales Ram ´on, and Y . Mezouar, “Visual- tactile fusion for 3d objects reconstruction from a single depth view and a single gripper touch for robotics tasks,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021,...

2021
[20]

Shapemap 3-d: Efficient shape mapping through dense touch and vision,

S. Suresh, Z. Si, J. G. Mangelson, W. Yuan, and M. Kaess, “Shapemap 3-d: Efficient shape mapping through dense touch and vision,” in2022 International Conference on Robotics and Automation (ICRA), 2022, pp. 7073–7080

2022
[21]

Active visuo-haptic object shape completion,

L. Rustler, J. Lundell, J. K. Behrens, V . Kyrki, and M. Hoffmann, “Active visuo-haptic object shape completion,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 5254–5261, 2022

2022
[22]

Active 3d shape reconstruction from vision and touch,

E. Smith, D. Meger, L. Pineda, R. Calandra, J. Malik, A. Romero Soriano, and M. Drozdzal, “Active 3d shape reconstruction from vision and touch,” inAdvances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, Eds., vo...

2021
[23]

3d shape reconstruction from vision and touch,

E. Smith, R. Calandra, A. Romero, G. Gkioxari, D. Meger, J. Malik, and M. Drozdzal, “3d shape reconstruction from vision and touch,” Advances in Neural Information Processing Systems, vol. 33, pp. 14 193–14 206, 2020

2020
[24]

Vihope: Visuotactile in-hand object 6d pose estimation with shape completion,

H. Li, S. Dikhale, S. Iba, and N. Jamali, “Vihope: Visuotactile in-hand object 6d pose estimation with shape completion,”IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 6963–6970, 2023

2023
[25]

Gaussian process implicit surfaces for shape estimation and grasping,

S. Dragiev, M. Toussaint, and M. Gienger, “Gaussian process implicit surfaces for shape estimation and grasping,” in2011 IEEE Interna- tional Conference on Robotics and Automation, 2011, pp. 2845–2850

2011
[26]

Gaussian process im- plicit surfaces,

O. Williams and A. Fitzgibbon, “Gaussian process im- plicit surfaces,” inGaussian Processes in Practice, April
[27]

Learning efficient haptic shape exploration with a rigid tactile sensor array,

S. Fleer, A. Moringen, R. L. Klatzky, and H. Ritter, “Learning efficient haptic shape exploration with a rigid tactile sensor array,” PLOS ONE, vol. 15, no. 1, pp. 1–22, 01 2020. [Online]. Available: https://doi.org/10.1371/journal.pone.0226880

2020 doi
[28]

Active tactile exploration using shape- dependent reinforcement learning,

S. Jiang and L. L. Wong, “Active tactile exploration using shape- dependent reinforcement learning,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 8995– 9002

2022
[29]

Actexplore: Active tactile explo- ration on unknown objects,

A.-H. Shahidzadeh, S. J. Yoo, P. Mantripragada, C. D. Singh, C. Ferm ¨uller, and Y . Aloimonos, “Actexplore: Active tactile explo- ration on unknown objects,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 3411–3418

2024
[30]

Active multi-contact contin- uous tactile exploration with gaussian process differential entropy,

D. Driess, D. Hennes, and M. Toussaint, “Active multi-contact contin- uous tactile exploration with gaussian process differential entropy,” in 2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 7844–7850

2019
[31]

T. Inc. Shokac chip 6dof-p18. Accessed on February 25, 2024. [Online]. Available: http://www.touchence.jp/en/company.html

2024
[32]

The ycb object and model set: Towards common benchmarks for manipulation research,

B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar, “The ycb object and model set: Towards common benchmarks for manipulation research,” in2015 International Conference on Advanced Robotics (ICAR), 2015, pp. 510–517

2015
[119]

3789–3799

PMLR, 13–18 Jul 2020, pp. 3789–3799. [Online]. Available: https://proceedings.mlr.press/v119/gropp20a.html

2020
[2007]

Available: https://www.microsoft.com/en-us/research/ publication/gaussian-process-implicit-surfaces-2/

[Online]. Available: https://www.microsoft.com/en-us/research/ publication/gaussian-process-implicit-surfaces-2/

Pith tools

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

[1] [1]

Humans integrate visual and haptic information in a statistically optimal fashion,

M. O. Ernst and M. S. Banks, “Humans integrate visual and haptic information in a statistically optimal fashion,”Nature, vol. 415, no. 6870, pp. 429–433, 2002

2002

[2] [2]

Deepsdf: Learning continuous signed distance functions for shape representation,

J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “Deepsdf: Learning continuous signed distance functions for shape representation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

work page 2019

[3] [3]

Shape completion enabled robotic grasping,

J. Varley, C. DeChant, A. Richardson, J. Ruales, and P. Allen, “Shape completion enabled robotic grasping,” in2017 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2017, pp. 2442–2447

work page 2017

[4] [4]

Robust grasp planning over uncertain shape completions,

J. Lundell, F. Verdoja, and V . Kyrki, “Robust grasp planning over uncertain shape completions,” in2019 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 1526–1532

work page 2019

[5] [5]

Completion and reconstruction with primitive shapes,

R. Schnabel, P. Degener, and R. Klein, “Completion and reconstruction with primitive shapes,”Computer Graphics F orum, vol. 28, no. 2, pp. 503–512, 2009. [Online]. Available: https://onlinelibrary.wiley.com/ doi/abs/10.1111/j.1467-8659.2009.01389.x

arXiv 2009

[6] [6]

Active haptic perception in robots: a review,

L. Seminara, P. Gastaldo, S. J. Watt, K. F. Valyear, F. Zuher, and F. Mastrogiovanni, “Active haptic perception in robots: a review,” Frontiers in neurorobotics, vol. 13, p. 53, 2019

work page 2019

[7] [7]

Active tactile exploration with uncer- tainty and travel cost for fast shape estimation of unknown objects,

T. Matsubara and K. Shibata, “Active tactile exploration with uncer- tainty and travel cost for fast shape estimation of unknown objects,” Robotics and Autonomous Systems, vol. 91, pp. 314–326, 2017

work page 2017

[8] [8]

Active perception: Building objects’ models using tactile exploration,

N. Jamali, C. Ciliberto, L. Rosasco, and L. Natale, “Active perception: Building objects’ models using tactile exploration,” in2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids), 2016, pp. 179–185

work page 2016