Pith. sign in

REVIEW 3 major objections 6 minor 88 references

Gaussian Process-Based Active Exploration Strategies in Vision and Touch

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The thesis argues that one Gaussian Process Distance Field can carry both visual and tactile observations, letting a robot actively refine an object's geometry by choosing camera poses and contact points that maximize information gain…

desk verdict A promising GPDF-based vision-touch exploration pipeline whose central claim is undercut by experiments that only validate the uncertainty proxies it optimizes, not reconstruction accuracy. read the letter →

arxiv 2507.05522 v1 pith:YUFUP32W submitted 2025-07-07 cs.RO

Ho Jin Choi , Nadia Figueroa This is my paper
classification cs.RO
keywords GaussianProcessDistanceFieldactiveperceptiontactileexplorationnext-best-viewplanningdifferentiablerenderingsurfaceuncertaintyvisuo-tactilefusionmaterialclassification
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 thesis sets out to show that one probabilistic representation, the Gaussian Process Distance Field (GPDF), can serve as the shared world model for active exploration with both vision and touch. Starting from a single RGB-D view, the robot builds a signed distance field from the point cloud, then chooses the next camera pose by estimated information gain and the next contact point by surface uncertainty, updating the same field with each observation. If this works, a robot can refine complex object geometry without large pretraining datasets or a neural shape model, because the field is constructed directly from sensor points and updated incrementally. The thesis additionally argues that the same field can hold color, semantic image features, and material labels, so the exploration loop extends beyond geometry.

What carries the argument

The load-bearing object is the Gaussian Process Distance Field: Gaussian Process regression over observed surface points whose y-values are fixed to one, combined with a reverting function that maps the occupancy posterior back to a distance so that the field satisfies $|\nabla d|=1$. The covariance kernel $\exp(-d/\ell)$ is singled out because its reverting function is bijective, which produces signed distance; ray marching then iteratively pulls query points to the surface. Surface uncertainty is derived either from a Mahalanobis distance between the kernel derivative and the inferred gradient, or from deviation of $\|\nabla d\|$ from one, and it is this uncertainty that drives touch contact selection through projected gradient ascent. Visual exploration is carried by differentiable volumetric rendering of the SDF, whose color and depth estimates give ensemble-based information gain for next-best-view choice; inducing points, voxel downsampling, and Nyström approximation keep the Gaussian Process cost manageable.

What would settle it

Measure the point-to-surface or Chamfer distance between the GPDF surface and a ground-truth scan of the same object after each exploration step; if information-gain-selected camera poses and uncertainty-selected contact points do not reduce that error faster than random selection, the paper's planning claim would be overturned.

Watch

Extended reading notes

Core claim

The central claim is that the Gaussian Process Distance Field is a unified, training-free representation for active visuo-tactile perception. By applying a Gaussian Process to surface points and using a reverting function that inverts the kernel, the model outputs signed distance, analytic gradient and Hessian, and surface uncertainty at any query point; with the exponential kernel $\exp(-d/\ell)$ the reverting function is bijective, giving a true signed distance field that ray marching can refine. The thesis shows this field can be rendered through differentiable volumetric ray marching, so color images with known poses can optimize the underlying point cloud even where depth is unreliable, and an ensemble of GPDFs with different interpolation lengths supplies the information-gain signal for choosing the next camera view. For touch, gradient ascent on surface uncertainty, projected onto the surface tangent plane, finds the most uncertain points to contact; real-robot experiments show maximum surface uncertainty decreasing as tactile exploration adds points. The same GP machinery is extended to RGB color, CLIP embeddings, and force-based material classification, making the claimed contribution a multi-modal active perception pipeline rather than a geometry-only method.

Load-bearing premise

The load-bearing premise is that the information-gain scores and surface-uncertainty values computed from the GPDF are faithful proxies for real reconstruction improvement; the thesis plans every next view and contact from these proxies but never checks them against ground-truth geometry.

Editorial extensions

If this is right

  • A robot can start from a single RGB-D view and refine a full distance field by choosing each next camera pose from information gain and each next contact from surface uncertainty, with no network training on large shape datasets.
  • New camera views and tactile contacts can be folded into the existing field via incremental block-matrix updates, so the representation supports online active perception rather than full re-optimization.
  • Because color, CLIP embeddings, and material features share the same GP machinery, exploration policies designed for geometry transfer to semantic and physical property estimation.
  • Differentiable rendering of the field makes it possible to recover geometry in reflective or translucent regions where the depth camera is noisy, by optimizing the point cloud against ordinary color images.
  • Scalable GP approximations (inducing points, voxel sampling, Nyström) bring the pipeline from cubic to roughly linear cost, permitting thousands of sensor points in the loop.

Reading between the lines

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

  • Editorial inference: a natural next experiment, left implicit in the thesis, is to compare information-gain-driven view selection against random view selection using Chamfer distance to a ground-truth scan; this would directly test whether the uncertainty proxies track true reconstruction improvement.
  • Editorial inference: because the field is incremental and explicit, the same pipeline could be extended to deformable objects or continuous re-grasping if the point cloud were updated online; the thesis assumes static rigid objects.
  • Editorial inference: the CLIP feature field opens a route to language-guided exploration, where the robot queries which surface region matches a word and plans camera or touch motions to reduce semantic uncertainty; the thesis stores and renders these features but does not run the semantic exploration loop.
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 / 6 minor

Summary. The thesis proposes a Gaussian Process Distance Field (GPDF) as a unified representation for active multimodal perception, combining dense vision (RGBD plus differentiable rendering) and sparse tactile measurements. It derives differentiable rendering and information-gain measures for camera-pose selection, surface-uncertainty-guided tactile exploration, incremental GP updates, and scalable kernel approximations. The pipeline is instantiated in simulation and on a Franka robot with a customized DIGIT sensor, and a separate rolling-based tactile material-classification study is presented. The central claim is that information-gain-driven exploration with this representation enables training-free, incrementally updatable, accurate 3D reconstruction of complex object geometries.

Significance. If substantiated, the GPDF-based framework would be a valuable contribution: it combines a training-free implicit representation with analytic gradients, uncertainty quantification, incremental updates, and multimodal fusion, and it demonstrates differentiable rendering from a GP distance field, an unusual and potentially useful capability. Strengths include the detailed treatment of kernel approximations (Section 3.5, Figs. 3-9 to 3-11), the custom tactile-sensor calibration pipeline (Sections 2.2-2.3), and real-robot integration. The main weakness is that the empirical validation of the exploration benefit is internal: the reported curves show the optimized information metrics decreasing, but no ground-truth reconstruction error, no repeated trials, and no comparison to baselines are provided. The material-classification results (Table 5.1) are also too weak to support the broader multimodal claims. The framework is plausible, but the current evidence does not yet establish the central claim.

major comments (3)
  1. [§4.1.2 and §4.2.1, Figs. 4-14 and 4-21] The quantitative evidence for the central claim that information-gain-guided exploration improves geometry is the decrease of the very objectives being optimized: Fig. 4-14 shows maximum information gain dropping faster under IG-guided view selection than random selection, and Fig. 4-21 shows maximum surface uncertainty falling as tactile contacts are added. Because the selection policy chooses views/contacts with high IG or high uncertainty, these decreases are expected even if the model's geometric accuracy does not improve. No ground-truth reconstruction metric (e.g., Chamfer distance to a high-fidelity scan, held-out depth error, or surface-normal error) is reported, there are no repeated trials or error bars, and no comparison to random contact selection or to existing GPIS-based tactile exploration methods such as those in Refs. [48,49,51]. Section 6.2 lists scalability and exploration-strategy limitations but does not acknowledge this absence of external geometric validation. This is the load-bearing gap: without such validation, the experiments cannot distinguish effective active exploration from a closed active-learning loop that simply reduces its own surrogate.
  2. [§5.2.1, Table 5.1] The material-classification experiments, which are the main support for the 'beyond geometry' claim, achieve cross-validation accuracies of 36.37% (SVM), 33.11% (MLP), and 29.87% (Random Forest) on seven materials (chance 14.29%), with training accuracies of 36.05%, 82.99%, and 79.59%, respectively. The large train-CV gap indicates severe overfitting, and the paper's own hypothesis that force amplitude alone is insufficient is stated but not followed by any experiments with the higher-dimensional features (raw images, deformation vectors) it suggests. Section 5.2.2 then proposes GP-based incorporation of classification into the multimodal field, but no experimental validation of that integration is provided. As presented, this chapter does not establish that tactile exploration can reliably identify materials, and the claim in Section 1.4 that the framework demonstrates potential for modeling surface properties beyond geometry is therefore premature.
  3. [§3.2.2 and §4.2.1, Eq. (3.4) vs Eq. (4.11)] The relationship between the two uncertainty quantities is never made explicit. Equation (3.4) defines a 'surrogate for distance uncertainty' using a Mahalanobis-like expression with a covariance Σ_ô(x) that is not defined in the text, and the notation ∂|k(d)|/∂d|_dhat is ambiguous (evaluated at what value, and which kernel derivative?). Later, Section 4.2.1 and Eq. (4.11) define tactile surface uncertainty in terms of the gradient of the GP predictive variance var(o(x)), which is a different quantity. Since Fig. 4-21 reports 'maximum surface uncertainty decreasing' without stating which of these measures is plotted, the reader cannot verify what the tactile exploration actually optimized or whether the reported decrease is the one promised by the surrogate. This needs to be clarified and, ideally, both quantities should be reported.
minor comments (6)
  1. [§1.4] The phrase 'Surface Uncertianty Quatification' contains two typos; it should read 'Surface Uncertainty Quantification.'
  2. [§3.2.3, Eq. (3.6)] The terms ∂f̄ and Δf̄ are not defined; the text says 'a gradient of one mean function value' and 'a gradient of n mean function values' but does not specify the differentiation variable or the exact construction of these quantities.
  3. [Appendix A] The table caption uses 'Generalized Probability Density Function (GPDF)'; this should be 'Gaussian Process Distance Field (GPDF).'
  4. [Fig. 4-14] The y-axis label gives no units for information gain, and no error bars are shown; please state in the caption how information gain is computed (number of ensemble members, interpolation values, etc.).
  5. [§5.2.1] The description of cross-validation ('dividing the dataset into three subsets') should state whether the splits are per-trial or per-sample; with 187 data points from 3 trials per material, the standard error is large and should be reported.
  6. [§4.1.2] The proposed ensemble of GPDFs with varying interpolation values is not compared against the alternative single-GPDF method mentioned in the same paragraph; a brief quantitative comparison would increase confidence in the choice.

Circularity Check

2 steps flagged · score 6.0 of 10

Exploration is validated by its own objectives: the tactile 'success' curve (max surface uncertainty decreasing) is mathematically forced by the GP posterior variance update, and the visual 'convergence' curve is the information-gain metric the planner maximizes, with no ground-truth geometry comparison.

  1. fitted input called prediction [Section 4.1.5 (Implementation); Figure 4-14]
    "In contrast to random view selection in simulation, employing information gain to guide the exploration accelerates the convergence of the GPDFs, as illustrated in Figure 4-14. Choose the next-best-view from a set of predefined pose candidates based on the information gain."

    The only quantitative evidence for 'convergence' is the drop in maximum information gain, yet that same IG is the objective the next-view rule maximizes ('choose the next-best-view ... based on the information gain'). IG (Eq. 4.4, adopted from He et al.) is evaluated with an ensemble of GPDFs fitted to the already-collected images; after a pose chosen because it maximizes IG is observed and the model is re-fit on that view, the candidate-pose IG falls simply because the model has absorbed that view, even if the underlying geometry is no more accurate. Figure 4-14's random-selection baseline is scored with the same self-referential metric, and no ground-truth reconstruction error is reported in Chapter 4.

  2. self definitional [Section 4.2.2 (Implementation); Figure 4-21]
    "Identify the most uncertain surface point based on the GP model. Move the robot to the identified point and add it as an observation point. ... From Figure 4-21, it is evident that as more points are explored using tactile sensing, the maximum surface uncertainty decreases."

    The claimed success measure, maximum surface uncertainty decreasing (Fig. 4-21), is exactly the function whose argmax defines the contact choice: Section 4.2.1 ascends the variance gradient (Eq. 4.11) and targets the most uncertain surface point (Eq. 4.14), which is then added as an observation point. For fixed hyperparameters, the GP posterior variance (Eqs. 3.2/3.6) never increases at any query location when a training point is added (positive-definite kernel with PSD noise augmentation), so the maximum surface uncertainty is forced to fall each iteration by the conditioning equations.

full rationale

The central empirical claims, that information-gain-guided views 'accelerate the convergence of the GPDFs' and that touching uncertain regions yields 'successful' exploration, rest only on plots of the very objectives the planner optimizes. The tactile result is the stronger case: contact points are chosen at the argmax of the GP's surface variance, and adding an observation there makes the posterior variance non-increasing at every point (Eqs. 3.2/3.6), so the monotone decrease in Figure 4-21 is a consequence of the conditioning equations rather than evidence of geometric improvement. The visual result is partly self-referential: maximum IG is both the selection objective and the convergence metric, and no ground-truth reconstruction error is measured anywhere in Chapter 4, nor is any comparison made against GPIS-based tactile exploration methods (Driess et al., Khadivar et al.). Notably, Section 6.2 flags scalability and discrete-sampling limitations but never acknowledges the absence of geometric validation of the exploration benefit, which is the load-bearing gap that leaves these self-referential metrics as the only quantitative support. The self-citation [10] (the authors' own prior work) for the signed-distance property of the Matern 1/2 kernel is not load-bearing: Table A.1 (sourced to [1,2]) gives the reverting function r(o) = -l*log(o), whose inversion of k(d) = exp(-d/l) is directly checkable in the appendix, independent of [10]. The remainder of the thesis, including the GPDF formulation from [1,22], kernel approximations, incremental updates, QP control, and the honestly reported low material-classification accuracies (Table 5.1, cross-validation 29-36%), is self-contained against external work and shows no equation-level circularity. The circularity is partial but attaches to the thesis's central claim, so the score is 6 rather than lower.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The framework rests on the GPDF formulation from prior work (Gentil et al. 2023, Wu et al. 2020), plus the hand-tuned parameters and modeling assumptions listed above. The most consequential unvalidated ingredient is the use of GP variance and eikonal deviation as surface uncertainty proxies, because exploration decisions and the main empirical plots depend on them without ground-truth verification. No new latent physical entity is introduced; the customized DIGIT elastomer is a hardware modification, not a postulated entity.

free parameters (8)
  • GP lengthscale l = Not reported; heuristic l >= 2x nominal point spacing
    Controls interpolation versus extrapolation in GPDF and strongly affects distance and uncertainty fields. Chosen by hand per experiment in Section 3.2.2 and Appendix C.
  • Sigmoid smoothness parameter s in volumetric rendering = Not reported; described as a high positive constant
    Appears in Eq. 4.3 and controls opacity sharpness along rendered rays. Not specified numerically.
  • Number of inducing points m = 1000 and 2000 in Figure 4-4
    Sets the trade-off between reconstruction detail and GP complexity. Chosen by hand.
  • Input or observation noise covariance = Not reported
    Used in noisy-input GPDF in Eq. 3.6. The text says it can be fixed from sensor specifications, but no values are given.
  • Tactile stiffness matrix K = Not reported; fit with gradient descent
    Converts marker deformation to contact force in Section 2.3.3. Fit against a Bota force-torque sensor with a sparsity constraint.
  • Inverse reflectance function matrix R^-1 = 2x11 matrix fit by incremental solver
    Maps tactile camera color gradients to depth gradients in Section 2.3.1. Calibrated using a 6.4mm indenter.
  • Classifier hyperparameters = SVM C=10^3, gamma=10^-4; MLP 64x64x64
    Grid-searched in Section 5.2.1 and likely contributes to the large gap between training and cross-validation accuracy.
  • Voxel downsampling size = e.g. 2 cm unit voxel
    Heuristic approximation in Section 3.4 that changes the number and uncertainty of GPDF input points.
assumptions (6)
  • standard math The Matérn 1/2 kernel reverting function r(o) = -l log(o) is bijective and yields a signed distance field.
    Section 3.2.2 and Appendix A rely on this bijectiveness to convert the GP occupancy field into an SDF. It is taken from prior GPDF work.
  • domain assumption Objects are static, rigid, and within the robot workspace, and the robot and sensor kinematics are known.
    Section 1.2 states these assumptions explicitly. All exploration and control depend on them.
  • domain assumption Each tactile contact can be modeled as a point on a rigid object surface, with all deformation confined to the sensor.
    Section 1.3.3 adopts the point-tactile model and ignores object deformation and finite contact patches.
  • ad hoc to paper GPDF predictive variance and the eikonal deviation |1 - ||grad d||| are valid measures of geometric surface uncertainty.
    These proxies drive exploration in Sections 3.2.2 and 4.2.1, but the thesis does not validate them against ground-truth geometry.
  • domain assumption Ensemble information gain computed from rendered color and depth approximates the true value of a camera pose.
    Section 4.1.2 adopts the method of He et al. 2023 without calibration or convergence analysis.
  • ad hoc to paper CLIP features can be spatially interpolated with GP kernels and retain semantic meaning.
    Section 4.1.3 visualizes semantic alignment, but no task-level evaluation is performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gaussian Process-Based Active Exploration Strategies in Vision and Touch." pith.science (2026). https://pith.science/paper/YUFUP32W

@misc{pith2026250705522,
  author       = {Pith},
  title        = {Pith review of: Gaussian Process-Based Active Exploration Strategies in Vision and Touch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YUFUP32W}},
  note         = {Machine review of arXiv:2507.05522}
}
read the original abstract

Robots struggle to understand object properties like shape, material, and semantics due to limited prior knowledge, hindering manipulation in unstructured environments. In contrast, humans learn these properties through interactive multi-sensor exploration. This work proposes fusing visual and tactile observations into a unified Gaussian Process Distance Field (GPDF) representation for active perception of object properties. While primarily focusing on geometry, this approach also demonstrates potential for modeling surface properties beyond geometry. The GPDF encodes signed distance using point cloud, analytic gradient and Hessian, and surface uncertainty estimates, which are attributes that common neural network shape representation lack. By utilizing a point cloud to construct a distance function, GPDF does not need extensive pretraining on large datasets and can incorporate observations by aggregation. Starting with an initial visual shape estimate, the framework iteratively refines the geometry by integrating dense vision measurements using differentiable rendering and tactile measurements at uncertain surface regions. By quantifying multi-sensor uncertainties, it plans exploratory motions to maximize information gain for recovering precise 3D structures. For the real-world robot experiment, we utilize the Franka Research 3 robot manipulator, which is fixed on a table and has a customized DIGIT tactile sensor and an Intel Realsense D435 RGBD camera mounted on the end-effector. In these experiments, the robot explores the shape and properties of objects assumed to be static and placed on the table. To improve scalability, we investigate approximation methods like inducing point method for Gaussian Processes. This probabilistic multi-modal fusion enables active exploration and mapping of complex object geometries, extending potentially beyond geometry.

Figures

Figures reproduced from arXiv: 2507.05522 by the authors.

Figure 1-1
Figure 1-1. Babies enhance cognitive abilities through active exploration, relying on [PITH_FULL_IMAGE:figures/full_fig_p013_1-1.png] view at source ↗
Figure 1-2
Figure 1-2. Active perception pipeline for different modalities and object properties [PITH_FULL_IMAGE:figures/full_fig_p014_1-2.png] view at source ↗
Figure 1-3
Figure 1-3. An object or scene, illustrated as a sphere, is a set, Ω, in n-dimensional [PITH_FULL_IMAGE:figures/full_fig_p015_1-3.png] view at source ↗
Figures from the paper (43 more)
Figure 1-5
Figure 1-5. Figure 1-5: Illustration of ray casting for NERF [4] rendering where ray comes out from the camera and discrete sam￾ples along the ray on the ray sequentially, defined as xti = r0 + rdti , where tmin ≤ t ≤ tmax and i is the index of points incrementally increasing, we obtain: αi…
Figure 1-6
Figure 1-6. Figure 1-6: Illustration of a real tactile model where deformation happens [PITH_FULL_IMAGE:figures/full_fig_p022_1-6.png]
Figure 1-8
Figure 1-8. Figure 1-8: The exploration begins with a rough estimate through vision, followed by [PITH_FULL_IMAGE:figures/full_fig_p024_1-8.png]
Figure 2-1
Figure 2-1. Figure 2-1: Hardware setup for the experiments. The left image provides an overview [PITH_FULL_IMAGE:figures/full_fig_p026_2-1.png]
Figure 2-2
Figure 2-2. Figure 2-2: Image from the original DIGIT paper [5]. Exploded view of a single [PITH_FULL_IMAGE:figures/full_fig_p027_2-2.png]
Figure 2-3
Figure 2-3. Figure 2-3: Manufacturing process of custom silicone elastomer with and without [PITH_FULL_IMAGE:figures/full_fig_p027_2-3.png]
Figure 2-4
Figure 2-4. Figure 2-4: Image processing pipeline for the DIGIT tactile sensor that works in 30-40 [PITH_FULL_IMAGE:figures/full_fig_p029_2-4.png]
Figure 2-5
Figure 2-5. Figure 2-5: Image from the original CPD paper [6] illustrating non-rigid point cloud [PITH_FULL_IMAGE:figures/full_fig_p031_2-5.png]
Figure 2-6
Figure 2-6. Figure 2-6: Markers and its neighbors forming hexahedron elements [PITH_FULL_IMAGE:figures/full_fig_p032_2-6.png]
Figure 2-7
Figure 2-7. Figure 2-7: Force calibration result structural relationships between markers. While this issue can be mitigated with a large amount of data, we opt for a different approach. We utilize gradient descent with constraining stiffness matrix values that do not align with neighboring…
Figure 3-1
Figure 3-1. Figure 3-1: Different representations of object shape. Image from [7] [PITH_FULL_IMAGE:figures/full_fig_p035_3-1.png]
Figure 3-3
Figure 3-3. Figure 3-3: Signed distance of a capsule. Image from [9]. [69, 70]. Superquadrics does not give SDF directly as in [PITH_FULL_IMAGE:figures/full_fig_p036_3-3.png]
Figure 3-4
Figure 3-4. Figure 3-4: An illustration of GPDF using a 3D point cloud. From a raw point cloud, [PITH_FULL_IMAGE:figures/full_fig_p038_3-4.png]
Figure 3-5
Figure 3-5. Figure 3-5: A 1D slice of a distance field (top: l = 1, bottom: l = 0.1) of a sphere point cloud. The image in the middle illustrates a comparison between the ground truth SDF and the initial estimation derived from the GPDF. On the right, the image illustrates an improved estim…
Figure 3-6
Figure 3-6. Figure 3-6: Original point cloud of Stanford Bunny (left) and downsampled point [PITH_FULL_IMAGE:figures/full_fig_p044_3-6.png]
Figure 3-7
Figure 3-7. Figure 3-7: Results of GPDF for raw input and down-sampled inputs. [PITH_FULL_IMAGE:figures/full_fig_p045_3-7.png]
Figure 3-8
Figure 3-8. Figure 3-8: An illustration of voxel down-sampling. From 100,000 points on the object [PITH_FULL_IMAGE:figures/full_fig_p047_3-8.png]
Figure 3-9
Figure 3-9. Figure 3-9: GPDF results (1-2 row: within the approximation bounding box, 3-4 row: [PITH_FULL_IMAGE:figures/full_fig_p052_3-9.png]
Figure 3-11
Figure 3-11. Figure 3-11: Inference time comparison between kernel approximation methods 3-11. In terms of computational speed, SKI is the fastest method for both train￾ing and inference, achieving less than 0.1 seconds for training and 2-3 milliseconds for inference. This is followed by the…
Figure 4-2
Figure 4-2. Figure 4-2: Illustration of volumetric ray marching original KL divergence needs the set of a point and its observation, this new method provides an additional avenue for enhancement. Depending on the situation, we can use differentiable rendering to optimize point clouds, or us…
Figure 4-3
Figure 4-3. Figure 4-3: Optimized GPDF using multi-view color images [PITH_FULL_IMAGE:figures/full_fig_p060_4-3.png]
Figure 4-4
Figure 4-4. Figure 4-4: Point cloud obtained from optimized GPDF (left: initialized with 1000 [PITH_FULL_IMAGE:figures/full_fig_p060_4-4.png]
Figure 4-5
Figure 4-5. Figure 4-5: Information gain from GPDF ensemble with varying interpolations in sim [PITH_FULL_IMAGE:figures/full_fig_p062_4-5.png]
Figure 4-6
Figure 4-6. Figure 4-6: Multimodal incorporation in Gaussian Process Distance Field [PITH_FULL_IMAGE:figures/full_fig_p063_4-6.png]
Figure 4-7
Figure 4-7. Figure 4-7: Simulation setup of YCB datasets on the table with 8 different camera [PITH_FULL_IMAGE:figures/full_fig_p064_4-7.png]
Figure 4-8
Figure 4-8. Figure 4-8: Rendered image of GPDF and CLIP features compared with the language [PITH_FULL_IMAGE:figures/full_fig_p065_4-8.png]
Figure 4-9
Figure 4-9. Figure 4-9: Control based on the Information Gain from exploration strategies [PITH_FULL_IMAGE:figures/full_fig_p066_4-9.png]
Figure 4-10
Figure 4-10. Figure 4-10: Sphere collision model of a robot from cuRoBo [11] [PITH_FULL_IMAGE:figures/full_fig_p067_4-10.png]
Figure 4-11
Figure 4-11. Figure 4-11: Visual exploration pipeline [PITH_FULL_IMAGE:figures/full_fig_p068_4-11.png]
Figure 4-12
Figure 4-12. Figure 4-12: Simulation environment with Franka robot [PITH_FULL_IMAGE:figures/full_fig_p068_4-12.png]
Figure 4-14
Figure 4-14. Figure 4-14: In the real robot environment, we do not observe an initially consistent [PITH_FULL_IMAGE:figures/full_fig_p069_4-14.png]
Figure 4-15
Figure 4-15. Figure 4-15: Rendered GPDF results with different point cloud initialization(column [PITH_FULL_IMAGE:figures/full_fig_p070_4-15.png]
Figure 4-16
Figure 4-16. Figure 4-16: Gradient ascent towards higher uncertainty area (red: observed points, [PITH_FULL_IMAGE:figures/full_fig_p072_4-16.png]
Figure 4-17
Figure 4-17. Figure 4-17: Optimization using accurate projection (left) and dual-goal optimization [PITH_FULL_IMAGE:figures/full_fig_p072_4-17.png]
Figure 4-18
Figure 4-18. Figure 4-18: Tactile exploration pipeline [PITH_FULL_IMAGE:figures/full_fig_p073_4-18.png]
Figure 4-19
Figure 4-19. Figure 4-19: Tactile exploration of uncertain surface areas [PITH_FULL_IMAGE:figures/full_fig_p073_4-19.png]
Figure 4-20
Figure 4-20. Figure 4-20: Point cloud of environment and explored uncertain surface point col￾ored in green [PITH_FULL_IMAGE:figures/full_fig_p074_4-20.png]
Figure 5-1
Figure 5-1. Figure 5-1: Exploration pipeline for material properties [PITH_FULL_IMAGE:figures/full_fig_p076_5-1.png]
Figure 5-2
Figure 5-2. Figure 5-2: Hand motions from [12]. Through tactile exploration, humans can gather [PITH_FULL_IMAGE:figures/full_fig_p077_5-2.png]
Figure 5-4
Figure 5-4. Figure 5-4: Sensor wearing out during sliding exploration destroying the reflective coating necessary for depth estimation, as shown in [PITH_FULL_IMAGE:figures/full_fig_p078_5-4.png]
Figure 5-5
Figure 5-5. Figure 5-5: Different materials with its tactile sensor reading [PITH_FULL_IMAGE:figures/full_fig_p079_5-5.png]
Figure 5-6
Figure 5-6. Figure 5-6: Real robot doing a rolling strategy [PITH_FULL_IMAGE:figures/full_fig_p079_5-6.png]
Figure 5-8
Figure 5-8. Figure 5-8: Rolling trials and its force amplitudes Training accuracy Cross validation accuracy SVM (RBF kernel) 36.05% 36.37% MLP (64x64x64) 82.99% 33.11% Random Forest 79.59% 29.87% [PITH_FULL_IMAGE:figures/full_fig_p080_5-8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 71 canonical work pages

  1. [1]

    Accurate Gaussian-Process-based Distance Fields with applications to Echolocation and Mapping

    C. L. Gentil, O.-L. Ouabi, L. Wu, C. Pradalier, and T. Vidal-Calleja, “Accurate gaussian process distance fields with applications to echolocation and mapping,” ArXiv, vol. abs/2302.13005, 2023

    work page Pith review arXiv 2023

  2. [2]

    Gaussian processes for machine learning,

    S. T. Ounpraseuth, “Gaussian processes for machine learning,” Journal of the American Statistical Association, vol. 103, pp. 429 – 429, 2008

    2008

  3. [3]

    The development of embodied cognition: Six lessons from babies,

    L. B. Smith and M. Gasser, “The development of embodied cognition: Six lessons from babies,” Artificial Life, vol. 11, pp. 13–29, 2005

    2005

  4. [4]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Commun. ACM, vol. 65, pp. 99–106, 2020

    2020

  5. [5]

    Digit: A novel design for a low-cost compact high-resolution tactile sensor with appli- cation to in-hand manipulation,

    M. Lambeta, P. wei Chou, S. Tian, B. Yang, B. Maloon, V. R. Most, D. Stroud, R. Santos, A. Byagowi, G. Kammerer, D. Jayaraman, and R. Calandra, “Digit: A novel design for a low-cost compact high-resolution tactile sensor with appli- cation to in-hand manipulation,” IEEE Robotics and Automation Letters, vol. 5, pp. 3838–3845, 2020

    2020

  6. [6]

    Point set registration: Coherent point drift,

    A. Myronenko and X. B. Song, “Point set registration: Coherent point drift,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 32, pp. 2262–2275, 2009

    2009

  7. [7]

    Advances in neural rendering,

    A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, Z. Xu, T. Simon, M. Nießner, E. Tretschk, L. Liu, B. Mildenhall, P. Srinivasan, R. Pandey, S. Orts-Escolano, S. Fanello, M. G. Guo, G. Wetzstein, J. y Zhu, C. Theobalt, M. Agrawala, D. B. Goldman, and M. Zollh¨ ofer, “Advances in neural rendering,” Computer Graphics Forum, vol. 41, 2021

    2021

  8. [8]

    Superquadrics and their geometric prop- erties,

    A. Jakliˇ c, A. Leonardis, and F. Solina, “Superquadrics and their geometric prop- erties,” 2000

    2000

Show all 88 references
  1. [9]

    Capsule - closest 3d

    iq, “Capsule - closest 3d.” https://www.shadertoy.com/view/stXXzl, Jul

  2. [10]

    Towards feasible dynamic grasping: Leveraging gaussian process distance field, se(3) equivariance and riemannian mixture mod- els,

    H. J. Choi and N. Figueroa, “Towards feasible dynamic grasping: Leveraging gaussian process distance field, se(3) equivariance and riemannian mixture mod- els,” ArXiv, vol. abs/2311.02576, 2023

    2023 arXiv

  3. [11]

    Curobo: Parallelized collision-free minimum-jerk robot motion genera- tion,

    B. Sundaralingam, S. K. S. Hari, A. Fishman, C. R. Garrett, K. V. Wyk, V. Blukis, A. Millane, H. Oleynikova, A. Handa, F. Ramos, N. Ratliff, and D. Fox, “Curobo: Parallelized collision-free minimum-jerk robot motion genera- tion,” ArXiv, vol. abs/2310.17274, 2023

    arXiv

  4. [12]

    Hand movements: A window into haptic object recognition,

    S. J. Lederman and R. L. Klatzky, “Hand movements: A window into haptic object recognition,” Cognitive Psychology, vol. 19, pp. 342–368, 1987

    1987

  5. [13]

    Revisiting active perception,

    R. Bajcsy, Y. Aloimonos, and J. K. Tsotsos, “Revisiting active perception,” CoRR, vol. abs/1603.02729, 2016

    2016 arXiv

  6. [14]

    Active perception,

    R. Bajcsy, “Active perception,” Proc. IEEE, vol. 76, pp. 966–1005, 1988

    1988

  7. [15]

    Object recognition using vision and touch,

    P. K. Allen and R. Bajcsy, “Object recognition using vision and touch,” in In- ternational Joint Conference on Artificial Intelligence , 1985

    1985

  8. [16]

    Morphing and sampling network for dense point cloud completion,

    M. Liu, L. Sheng, S. Yang, J. Shao, and S. Hu, “Morphing and sampling network for dense point cloud completion,” CoRR, vol. abs/1912.00280, 2019

    1912 arXiv

  9. [17]

    Foldingnet: Interpretable unsupervised learning on 3d point clouds,

    Y. Yang, C. Feng, Y. Shen, and D. Tian, “Foldingnet: Interpretable unsupervised learning on 3d point clouds,” CoRR, vol. abs/1712.07262, 2017

    2017 arXiv

  10. [18]

    PCN: point completion network,

    W. Yuan, T. Khot, D. Held, C. Mertz, and M. Hebert, “PCN: point completion network,” CoRR, vol. abs/1808.00671, 2018

    2018 arXiv

  11. [19]

    isdf: Real-time neural signed distance fields for robot percep- tion,

    J. Ortiz, A. Clegg, J. Dong, E. Sucar, D. Novotn´ y, M. Zollhoefer, and M. Mukadam, “isdf: Real-time neural signed distance fields for robot percep- tion,” ArXiv, vol. abs/2204.02296, 2022

    2022 arXiv

  12. [20]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,” ACM Transactions on Graphics (TOG) , vol. 42, pp. 1 – 14, 2023

    2023

  13. [21]

    Hand-eye calibration using dual quaternions,

    K. Daniilidis, “Hand-eye calibration using dual quaternions,” The International Journal of Robotics Research, vol. 18, pp. 286 – 298, 1999

    1999

  14. [22]

    Faithful euclidean distance field from log-gaussian process implicit surfaces,

    L. Wu, K. M. B. Lee, L. Liu, and T. A. Vidal-Calleja, “Faithful euclidean distance field from log-gaussian process implicit surfaces,” CoRR, vol. abs/2010.11487, 2020

    2010 arXiv

  15. [23]

    Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields,

    S. Lee, C. Le, W. Jiahao, A. Liniger, S. Kumar, and F. Yu, “Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields,” IEEE Robotics and Automation Letters , 2022. 85

    2022

  16. [24]

    Activenerf: Learning where to see with uncertainty estimation,

    X. Pan, Z. Lai, S. Song, and G. Huang, “Activenerf: Learning where to see with uncertainty estimation,” in European Conference on Computer Vision , 2022

    2022

  17. [25]

    Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering,

    L. Jin, X. Chen, J. Ruckin, and M. Popovi’c, “Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering,” 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 11305– 11312, 2023

    2023

  18. [26]

    Activermap: Radiance field for active mapping and planning,

    H. Zhan, J. Zheng, Y. Xu, I. D. Reid, and H. Rezatofighi, “Activermap: Radiance field for active mapping and planning,” ArXiv, vol. abs/2211.12656, 2022

    2022 arXiv

  19. [27]

    Fisherrf: Active view selection and un- certainty quantification for radiance fields using fisher information,

    W. Jiang, B. Lei, and K. Daniilidis, “Fisherrf: Active view selection and un- certainty quantification for radiance fields using fisher information,” ArXiv, vol. abs/2311.17874, 2023

    2023 arXiv

  20. [28]

    Active perception using neural radiance fields,

    S. He, C. D. Hsu, D. Ong, Y. Shao, and P. Chaudhari, “Active perception using neural radiance fields,” ArXiv, vol. abs/2310.09892, 2023

    2023 arXiv

  21. [29]

    Active perception for ambiguous objects classification,

    E. Safronov, N. A. Piga, M. Colledanchise, and L. Natale, “Active perception for ambiguous objects classification,” CoRR, vol. abs/2108.00737, 2021

    2021 arXiv

  22. [30]

    Segmentation of cluttered scenes through interactive perception,

    K. Hausman, C. Bersch, D. Pangercic, S. Osentoski, Z.-C. M´ arton, and M. Beetz, “Segmentation of cluttered scenes through interactive perception,” 2012

    2012

  23. [31]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sas- try, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning , 2021

    2021

  24. [32]

    Tactile tool manip- ulation /author=shirai, yuki; jha, devesh k.; raghunathan, arvind; hong, dennis /creationdate=may 2, 2023 /subject=robotics,

    Y. Shirai, D. K. Jha, A. U. Raghunathan, and D. W. Hong, “Tactile tool manip- ulation /author=shirai, yuki; jha, devesh k.; raghunathan, arvind; hong, dennis /creationdate=may 2, 2023 /subject=robotics,” 2023

    2023

  25. [33]

    Tactile dexterity: Manip- ulation primitives with tactile feedback,

    F. R. Hogan, J. Ballester, S. Dong, and A. Rodriguez, “Tactile dexterity: Manip- ulation primitives with tactile feedback,” 2020 IEEE International Conference on Robotics and Automation (ICRA) , pp. 8863–8869, 2020

    2020

  26. [34]

    Ca- ble manipulation with a tactile-reactive gripper,

    Y. She, S. Wang, S. Dong, N. Sunil, A. Rodriguez, and E. H. Adelson, “Ca- ble manipulation with a tactile-reactive gripper,” The International Journal of Robotics Research, vol. 40, pp. 1385 – 1401, 2019

    2019

  27. [35]

    Tactile perception of roughness and hardness to discriminate materials by friction-induced vibration,

    S. Ding, Y. Pan, M. Tong, and X. Zhao, “Tactile perception of roughness and hardness to discriminate materials by friction-induced vibration,” Sensors (Basel, Switzerland) , vol. 17, 2017. 86

    2017

  28. [36]

    Estimating object hardness with a gelsight touch sensor,

    W. Yuan, M. A. Srinivasan, and E. H. Adelson, “Estimating object hardness with a gelsight touch sensor,” 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 208–215, 2016

    2016

  29. [37]

    Multimodal zero-shot learning for tactile texture recognition,

    G. Cao, J. Jiang, D. Bollegala, M. Li, and S. Luo, “Multimodal zero-shot learning for tactile texture recognition,” ArXiv, vol. abs/2306.12705, 2023

    2023 arXiv

  30. [38]

    Few-shot learning of force-based motions from demonstration through pre-training of haptic represen- tation,

    M. Y. Aoyama, J. Moura, N. Saito, and S. Vijayakumar, “Few-shot learning of force-based motions from demonstration through pre-training of haptic represen- tation,” ArXiv, vol. abs/2309.04640, 2023

    2023 arXiv

  31. [39]

    Tactile sensing in dexterous robot hands - review,

    Z. Kappasov, J. A. Corrales, and V. Perdereau, “Tactile sensing in dexterous robot hands - review,” Robotics Auton. Syst., vol. 74, pp. 195–220, 2015

    2015

  32. [40]

    Soft tactile sensing skins for robotics,

    P. Roberts, M. Zadan, and C. Majidi, “Soft tactile sensing skins for robotics,” Current Robotics Reports, vol. 2, pp. 343 – 354, 2021

    2021

  33. [41]

    Gelsight: High-resolution robot tactile sensors for estimating geometry and force,

    W. Yuan, S. Dong, and E. H. Adelson, “Gelsight: High-resolution robot tactile sensors for estimating geometry and force,” Sensors (Basel, Switzerland), vol. 17, 2017

    2017

  34. [42]

    Gelslim 3.0: High-resolution mea- surement of shape, force and slip in a compact tactile-sensing finger,

    I. H. Taylor, S. Dong, and A. Rodriguez, “Gelslim 3.0: High-resolution mea- surement of shape, force and slip in a compact tactile-sensing finger,” 2022 In- ternational Conference on Robotics and Automation (ICRA) , pp. 10781–10787, 2021

    2022

  35. [43]

    Soft biomimetic optical tactile sensing with the tactip: A review,

    N. F. Lepora, “Soft biomimetic optical tactile sensing with the tactip: A review,” IEEE Sensors Journal , vol. 21, pp. 21131–21143, 2021

    2021

  36. [44]

    Multimodal proximity and visuotactile sensing with a selectively transmissive soft membrane,

    J. Yin, G. M. Campbell, J. H. Pikul, and M. H. Yim, “Multimodal proximity and visuotactile sensing with a selectively transmissive soft membrane,” 2022 IEEE 5th International Conference on Soft Robotics (RoboSoft) , pp. 802–808, 2022

    2022

  37. [45]

    Tactile sensors

    XELA Robotics, “Tactile sensors.” https://www.xelarobotics.com/tactile- sensors, 2018. [Online; accessed 9-April-2024]

    2018

  38. [46]

    Tactile sensors

    Contactile, “Tactile sensors.” https://contactile.com/products/#tactile_ sensors, 2019. [Online; accessed 9-April-2024]

    2019

  39. [47]

    Velocity level approximation of pressure field contact patches,

    J. G. Masterjohn, D. Guoy, J. Shepherd, and A. M. Castro, “Velocity level approximation of pressure field contact patches,”IEEE Robotics and Automation Letters, vol. 7, pp. 11593–11600, 2021

    2021

  40. [48]

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

    D. Driess, D. Hennes, and M. Toussaint, “Active multi-contact continuous tactile exploration with gaussian process differential entropy,” in 2019 International Conference on Robotics and Automation (ICRA) , pp. 7844–7850, May 2019. 87

    2019

  41. [49]

    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

  42. [50]

    Simultane- ous tactile exploration and grasp refinement for unknown objects,

    C. de Farias, N. Marturi, R. Stolkin, and Y. Bekiroglu, “Simultane- ous tactile exploration and grasp refinement for unknown objects,” CoRR, vol. abs/2103.00655, 2021

    2021 arXiv

  43. [51]

    Efficient shape mapping through dense touch and vision,

    S. Suresh, Z. Si, J. G. Mangelson, W. Yuan, and M. Kaess, “Efficient shape mapping through dense touch and vision,” CoRR, vol. abs/2109.09884, 2021

    2021 arXiv

  44. [52]

    Tactile SLAM: real-time inference of shape and pose from planar pushing,

    S. Suresh, M. Bauz´ a, K. Yu, J. G. Mangelson, A. Rodriguez, and M. Kaess, “Tactile SLAM: real-time inference of shape and pose from planar pushing,” CoRR, vol. abs/2011.07044, 2020

    2011 arXiv

  45. [53]

    Neural feels with neural fields: Visuo-tactile perception for in-hand manipulation,

    S. Suresh, H. Qi, T. Wu, T. Fan, L. Pineda, M. Lambeta, J. Malik, M. Kalakr- ishnan, R. Calandra, M. Kaess, J. Ortiz, and M. Mukadam, “Neural feels with neural fields: Visuo-tactile perception for in-hand manipulation,” 2023

    2023

  46. [54]

    Touch-gs: Visual-tactile supervised 3d gaussian splatting,

    A. Swann, M. Strong, W. K. Do, G. S. Camps, M. Schwager, and M. Kennedy, “Touch-gs: Visual-tactile supervised 3d gaussian splatting,” 2024

    2024

  47. [55]

    Active perception and modeling of deformable surfaces using gaussian processes and position-based dy- namics,

    S. Caccamo, P. G¨ uler, H. Kjellstr¨ om, and D. Kragic, “Active perception and modeling of deformable surfaces using gaussian processes and position-based dy- namics,” 2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids), pp. 530–537, 2016

    2016

  48. [56]

    Neus: Learn- ing neural implicit surfaces by volume rendering for multi-view reconstruction,

    P. Wang, L. Liu, Y. Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: Learn- ing neural implicit surfaces by volume rendering for multi-view reconstruction,” ArXiv, vol. abs/2106.10689, 2021

    2021 arXiv

  49. [57]

    Analysis and noise mod- eling of the intel realsense d435 for mobile robots,

    M. S. Ahn, H. Chae, D. Noh, H. Nam, and D. Hong, “Analysis and noise mod- eling of the intel realsense d435 for mobile robots,” in 2019 16th International Conference on Ubiquitous Robots (UR) , pp. 707–711, 2019

    2019

  50. [58]

    A new technique for fully autonomous and efficient 3d robotics hand/eye calibration,

    R. Y. Tsai and R. K. Lenz, “A new technique for fully autonomous and efficient 3d robotics hand/eye calibration,” IEEE Trans. Robotics Autom., vol. 5, pp. 345– 358, 1988

    1988

  51. [59]

    Products

    Smooth-On, “Products.” https://www.smooth-on.com/products/. [Online; accessed 9-April-2024]

    2024

  52. [60]

    Allsight: A low-cost and high-resolution round tactile sensor with zero-shot learning capability,

    O. Azulay, N. Curtis, R. Sokolovsky, G. Levitski, D. Slomovik, G. Lilling, and A. Sintov, “Allsight: A low-cost and high-resolution round tactile sensor with zero-shot learning capability,” IEEE Robotics and Automation Letters , vol. 9, pp. 483–490, 2023. 88

    2023

  53. [61]

    The hungarian method for the assignment problem,

    H. W. Kuhn, “The hungarian method for the assignment problem,” Naval Re- search Logistics (NRL), vol. 52, 1955

    1955

  54. [62]

    A mathematical analysis of the motion co- herence theory,

    A. L. Yuille and N. M. Grzywacz, “A mathematical analysis of the motion co- herence theory,” International Journal of Computer Vision , vol. 3, pp. 155–175, 1989

    1989

  55. [63]

    Towards a passive adaptive planar foot with ground orientation and contact force sensing for legged robots,

    R. Kaslin, H. Kolvenbach, L. Paez, K. Lika, and M. Hutter, “Towards a passive adaptive planar foot with ground orientation and contact force sensing for legged robots,” 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2707–2714, 2018

    2018

  56. [64]

    Differentiable collision detection for a set of convex primitives,

    K. Tracy, T. A. Howell, and Z. Manchester, “Differentiable collision detection for a set of convex primitives,” 2023 IEEE International Conference on Robotics and Automation (ICRA) , pp. 3663–3670, 2022

    2023

  57. [65]

    Factory: Fast contact for robotic assembly,

    Y. S. Narang, K. Storey, I. Akinola, M. Macklin, P. Reist, L. Wawrzyniak, Y. Guo, ´A. Morav´ anszky, G. State, M. Lu, A. Handa, and D. Fox, “Factory: Fast contact for robotic assembly,” ArXiv, vol. abs/2205.03532, 2022

    2022 arXiv

  58. [66]

    Representing robot geometry as distance fields: Applications to whole-body manipulation,

    Y. Li, Y. Zhang, A. Razmjoo, and S. Calinon, “Representing robot geometry as distance fields: Applications to whole-body manipulation,” 2023

    2023

  59. [67]

    Neural joint space implicit signed distance functions for reactive robot manipulator control,

    M. Koptev, N. Figueroa, and A. Billard, “Neural joint space implicit signed distance functions for reactive robot manipulator control,” IEEE Robotics and Automation Letters, vol. 8, pp. 480–487, 2023

    2023

  60. [68]

    Continuous implicit sdf based any-shape robot trajectory optimization,

    T. Zhang, J. Wang, C. Xu, A. Gao, and F. Gao, “Continuous implicit sdf based any-shape robot trajectory optimization,” 2023 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS) , pp. 282–289, 2023

    2023

  61. [69]

    Three dimensional object representation revisited,

    R. Bajcsy and F. Solina, “Three dimensional object representation revisited,” 1987

    1987

  62. [70]

    Robust and accurate su- perquadric recovery: a probabilistic approach,

    W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian, “Robust and accurate su- perquadric recovery: a probabilistic approach,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 2666–2675, 2021

    2022

  63. [71]

    Learn- ing continuous 3d reconstructions for geometrically aware grasping,

    M. V. der Merwe, Q. Lu, B. Sundaralingam, M. Matak, and T. Hermans, “Learn- ing continuous 3d reconstructions for geometrically aware grasping,” 2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 11516–11522, 2019

    2020

  64. [72]

    Gensdf: Two-stage learning of generaliz- able signed distance functions,

    G. Chou, I. Chugunov, and F. Heide, “Gensdf: Two-stage learning of generaliz- able signed distance functions,” ArXiv, vol. abs/2206.02780, 2022. 89

    2022 arXiv

  65. [73]

    Neural unsigned distance fields for im- plicit function learning,

    J. Chibane, A. Mir, and G. Pons-Moll, “Neural unsigned distance fields for im- plicit function learning,” ArXiv, vol. abs/2010.13938, 2020

    2010 arXiv

  66. [74]

    Gaussian processes for machine learning,

    M. W. Seeger, “Gaussian processes for machine learning,” International journal of neural systems , vol. 14 2, pp. 69–106, 2004

    2004

  67. [75]

    Gaussian process implicit surfaces,

    O. Williams and A. Fitzgibbon, “Gaussian process implicit surfaces,” inGaussian Processes in Practice, April 2007

    2007

  68. [76]

    Gaussian process training with input noise,

    A. Mchutchon and C. Rasmussen, “Gaussian process training with input noise,” in Advances in Neural Information Processing Systems (J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger, eds.), vol. 24, Curran Associates, Inc., 2011

    2011

  69. [77]

    Kernel inter- polation for scalable online gaussian processes,

    S. Stanton, W. J. Maddox, I. A. Delbridge, and A. G. Wilson, “Kernel inter- polation for scalable online gaussian processes,” in International Conference on Artificial Intelligence and Statistics , 2021

    2021

  70. [78]

    Practical hilbert space approximate bayesian gaussian processes for probabilistic programming,

    G. Riutort-Mayol, P.-C. Burkner, M. R. Andersen, A. Solin, and A. Vehtari, “Practical hilbert space approximate bayesian gaussian processes for probabilistic programming,” Statistics and Computing , vol. 33, 2020

    2020

  71. [79]

    Uni-fusion: Universal continuous mapping,

    Y. Yuan and A. Nuechter, “Uni-fusion: Universal continuous mapping,” ArXiv, vol. abs/2303.12678, 2023

    2023 arXiv

  72. [80]

    Segment tracing using local lipschitz bounds,

    E. Galin, ´E. Gu´ erin, A. Paris, and A. Peytavie, “Segment tracing using local lipschitz bounds,” Computer Graphics Forum, vol. 39, 2020

    2020

  73. [81]

    Forward inclusion functions for ray-tracing implicit surfaces,

    M. Aydinlilar and C. Zanni, “Forward inclusion functions for ray-tracing implicit surfaces,” Comput. Graph., vol. 114, pp. 190–200, 2023

    2023

  74. [82]

    Distilled feature fields enable few-shot language-guided manipulation,

    B. W. Shen, G. Yang, A. Yu, J. R. Wong, L. P. Kaelbling, and P. Isola, “Distilled feature fields enable few-shot language-guided manipulation,” ArXiv, vol. abs/2308.07931, 2023

    2023 arXiv

  75. [83]

    Yale-cmu-berkeley dataset for robotic manipulation re- search,

    B. C ¸ alli, A. Singh, J. Bruce, A. Walsman, K. Konolige, S. S. Srinivasa, P. Abbeel, and A. M. Dollar, “Yale-cmu-berkeley dataset for robotic manipulation re- search,” The International Journal of Robotics Research , vol. 36, pp. 261 – 268, 2017

    2017

  76. [84]

    A unified framework for co- ordinated multi-arm motion planning,

    S. S. M. Salehian, N. Figueroa, and A. Billard, “A unified framework for co- ordinated multi-arm motion planning,” The International Journal of Robotics Research, vol. 37, pp. 1205 – 1232, 2018

    2018

  77. [85]

    Learning to feel textures: Predicting perceptual similarities from unconstrained finger- surface interactions,

    B. A. Richardson, Y. Vardar, C. Wallraven, and K. J. Kuchenbecker, “Learning to feel textures: Predicting perceptual similarities from unconstrained finger- surface interactions,” IEEE Transactions on Haptics, vol. 15, pp. 705–717, 2022. 90

    2022

  78. [86]

    Robust surface recogni- tion with the maximum mean discrepancy: Degrading haptic-auditory signals through bandwidth and noise,

    B. Khojasteh, Y. Shao, and K. J. Kuchenbecker, “Robust surface recogni- tion with the maximum mean discrepancy: Degrading haptic-auditory signals through bandwidth and noise,” IEEE Transactions on Haptics , vol. 17, pp. 58– 65, 2024

    2024

  79. [87]

    Matern gaus- sian processes on riemannian manifolds,

    V. Borovitskiy, A. Terenin, P. Mostowsky, and M. P. Deisenroth, “Matern gaus- sian processes on riemannian manifolds,” ArXiv, vol. abs/2006.10160, 2020. 91

    2006 arXiv

  80. [2021]

    [Online; accessed 9-April-2024]. 84

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.