Pith. sign in

REVIEW 3 major objections 5 minor 27 references

A real-time RGB-D pipeline can generate feasible rock-breaking poses for autonomous impact hammers, with a 72% stable-contact success rate in scaled real-world trials.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 09:29 UTC pith:WLEOHSOC

load-bearing objection Solid integration paper with a real hardware test, but the headline 72% is a stable-contact rate, not a rock-fracture rate, and the conclusion overstates it. the 3 major comments →

arxiv 2607.20748 v1 pith:WLEOHSOC submitted 2026-07-22 cs.RO cs.CV

A real-time RGB-D perception pipeline for autonomous impact hammers in mining: self-filtering, rock segmentation and rock-breaking poses generation

Mart\'in Gallegos , Francisco Leiva , Patricio Loncomilla , Michelle Cort\'es , Javier Ruiz-del-Solar This is my paper
classification cs.RO cs.CV
keywords RGB-D perceptionimpact hammer automationrock-breaking pose generationsurface normal analysisreal-time perceptionpoint cloud segmentationmining roboticssecondary reduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the perception side of automating hydraulic impact hammers—machines that break oversized boulders on steel grates above ore passes—can be solved in real time. It presents a pipeline that takes RGB-D images from two ceiling-mounted stereo cameras and the hammer's kinematic state, removes the hammer's own geometry from the scene, segments rocks, and outputs prioritized target poses where the hammer should strike. The system runs at about 10 Hz with roughly 675 ms end-to-end latency on embedded hardware, and in a scaled real-world setup, 72% of pressing trials achieved stable chisel contact. If these numbers hold at full scale, the bottleneck of teleoperated secondary reduction in mining could be removed.

Core claim

The central claim is that surface-normal analysis over segmented rock point clouds, combined with the kinematic and operational constraints of a four-degree-of-freedom impact hammer, produces rock-breaking poses that are operationally feasible. Rather than placing one target at each rock's centroid, the method clusters points with similar surface orientations, defines a proto pose at the point closest to each cluster's centroid with orientation equal to the mean normal, and then projects that normal onto the plane defined by the target position and the hammer's rotation axis to obtain a reachable pitch angle. Thresholds discard poses that would strike too obliquely. In real-world pressing tr

What carries the argument

The load-bearing mechanism is the surface-normal-based pose generator. For each segmented rock point cloud, it computes local surface normals, clusters points with similar orientations via density-based clustering, forms a proto pose at the point closest to each cluster's centroid with orientation equal to the mean normal, projects that normal onto the t-plane defined by the target position and the hammer's rotation center to fix the pitch angle, and then applies soft and hard pitch-angle filters to keep only near-vertical, reachable strikes. This encodes the operational rule 'strike flat surfaces nearly perpendicular to the grizzly' directly into the pose geometry, which the paper credits f

Load-bearing premise

The key assumption is that a laboratory-scale scene—a mini-excavator fixed beside a 1.5-meter steel grate with two ceiling cameras—reproduces the conditions of a real underground ore-pass, including dust, lighting, full-size grizzlies, and machine dynamics; the 10 Hz throughput and 72% success rate were measured in that scaled setup.

What would settle it

A full-scale field trial at an actual ore pass, running the same pipeline on embedded hardware, in which the proportion of target poses that achieve stable chisel contact is recorded over a few hundred attempts; if the success rate falls well below 72% or the 10 Hz update cannot be sustained in dust and variable lighting, the paper's suitability claim is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The pipeline can be paired with a reaching controller and high-level state machines to close the loop: the hammer returns to a home configuration, selects the highest-priority pose, presses, and repeats, as demonstrated in the real-world validation.
  • At roughly 10 Hz with about 675 ms end-to-end latency, the system provides a continuous stream of target poses that is compatible with responsive closed-loop control on embedded hardware, without offline processing.
  • The robot-free 3D workspace representation, produced alongside the poses and with occlusions filled by the background model, can be consumed by downstream modules such as collision avoidance and motion planning.
  • Because poses are derived from local surface geometry rather than rock instance identity, imperfect segmentation does not block pose generation; it mainly perturbs the prioritization of poses, for example through rock-volume estimates.
  • The method is intentionally conservative: rocks with highly inclined surfaces may receive no pose at all, a design choice the paper argues is safer than proposing a slippage-prone target.
  • This is a field-tested extension: prior perception systems for rock breaking either used simple centroid targets or did not jointly enforce hammer kinematics and surface geometry, and this paper's contribution is precisely that joint enforcement running in real time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A stronger end-to-end test would fire the hammer rather than press: stable contact under sustained pressure does not guarantee effective energy transfer during an actual impact, so an impact trial could change the 72% figure.
  • The robot-free 3D workspace representation is a natural substrate for non-prehensile manipulation planning—repositioning unstable boulders by pushing—which the paper mentions only as future work.
  • The score weights and angular thresholds are hand-tuned for this setup; an adaptive or learned prioritization that maps geometric features to breakage outcomes could improve generalization across mines and rock types.
  • Because the pipeline assumes a quasi-static scene and ceiling-mounted cameras, transferring to a moving or differently placed sensor would require the optional point-cloud-based self-filter, whose latency was not measured separately in the paper.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a real-time RGB-D perception pipeline for autonomous hydraulic impact hammers in mining. The pipeline takes synchronized RGB-D frames from two ceiling-mounted stereo cameras plus the hammer's kinematic state, performs self-filtering of the robot from depth maps using approximated link meshes and modeled hydraulic cylinders, fills occluded regions via a depth-based background model, and then segments rocks by two complementary branches: RTMDet instance segmentation in RGB images and geometric point-cloud filtering/clustering. The outputs are prioritized rock-breaking target poses and a robot-free 3D workspace point cloud. The system is implemented on a Jetson AGX Orin using JAX, Open3D, and PyTorch, and is evaluated on a scaled setup consisting of a Bobcat E10 mini-excavator with a hydraulic impact hammer operating over a 1.5x1.5 m steel grate. Reported performance includes roughly 10 Hz output and about 675 ms end-to-end latency for pose generation, quantitative segmentation results for 1–7 rocks, and a real-world closed-loop pressing test in which 72% of 111 reachable target poses resulted in stable contact and a simulated strike via sustained pressure.

Significance. If the claims are upheld, the paper makes a useful systems contribution: it integrates self-filtering, occlusion management, image- and point-cloud-based rock segmentation, surface-normal-based pose generation, and prioritized pose selection into one embedded, real-time pipeline, with a real-world closed-loop validation. The paper is unusually transparent about parameter settings, hardware constraints, and the limitations of the experimental proxy. The main source of concern is not internal inconsistency but the strength of the headline conclusion: the closed-loop experiment measures stable pressing, not actual hammer impacts, so the transition from 'stable contact under sustained pressure' to 'practical applicability for autonomous impact hammer operation' needs either additional experimental evidence or a substantial re-framing.

major comments (3)
  1. [Section V-C1, Table II; Section VI; Abstract] The central real-world success metric is defined as stable chisel–rock contact while the boom presses with constant current for 5 seconds; the authors explicitly state that this is a 'simulated strike (via sustained pressure exertion)' and that the result 'does not imply that actuating the impact hammer’s end-effector in these situations would result in rock fractures.' Yet the abstract concludes that the system 'is suitable for real-time autonomous impact hammer operation,' and the conclusion states that the 72% success rate 'suggest[s] its practical applicability for autonomous impact hammer operation.' The load-bearing outcome variable is therefore a proxy, and the actual impact dynamics—vibration, rebound, higher slip forces, rock displacement—are never exercised. I would ask for either (i) an additional experiment with real hammer actuation, or (ii) a careful re-scoping of the title
  2. [Section V-C1, Table II] The 72% success rate is computed after excluding 6 of 117 poses because the reaching controller could not reach them. Since reachability is part of the overall objective ('operationally feasible rock-breaking poses'), the paper should report both the conditional rate (80/111 = 72.07%) and the intention-to-treat rate (80/117 = 68.4%), or at least state clearly that the headline number is conditional on successful reachability. In addition, no confidence interval is given; for 80 successes out of 111 trials, the Wilson 95% interval is approximately 63–80%, so the 72% figure has meaningful uncertainty that should be disclosed.
  3. [Section III, Section VI, Abstract] All experiments were performed in a laboratory-scale scene: a Bobcat E10 beside a 1.5x1.5 m steel grate with two ceiling-mounted ZED X cameras. The abstract's claim of suitability for 'autonomous impact hammer operation' extrapolates from this scaled setup to full-size underground ore-pass conditions. Differences in grizzly scale, camera placement, dust, lighting, and machine dynamics are not evaluated. This is not an internal inconsistency, but it is a correctness risk for the generalization implied by the headline claim. The paper should either temper the conclusion to the demonstrated scaled setup or add a discussion of which components are expected to transfer and which may require re-validation.
minor comments (5)
  1. [Section III] Typo: 'the problem consist of' should read 'the problem consists of.'
  2. [Section V-C1] Missing space in 'across11 different configurations'; should be 'across 11 different configurations.'
  3. [Appendix D, Table V] The RTMDet node is reported at approximately 15 Hz, but the appendix notes that this is achieved 'artificially' by reusing the most recent inference result. This is important context and should be restated in the main text near the 10 Hz claim, so the reader understands that the 10 Hz pose output includes repeated segmentation results.
  4. [Table II] The discrepancy between the quantitative 'No contact detected' (27.93%) and qualitative 'No contact detected' (24.32%) is explained only in prose. A brief note in the table caption that the quantitative/qualitative columns use different contact-detection definitions would improve readability.
  5. [Appendix C, Fig. 20] The architecture diagram is dense and some node names in the text ('pcl_vis_rock_segmentator') do not exactly match the figure labels. Aligning the names would reduce confusion.

Circularity Check

0 steps flagged

No significant circularity: the central result is an external physical measurement; self-citations are not load-bearing.

full rationale

The paper's derivation chain is empirical rather than a formal derivation. The headline 72% success rate is measured in real-world pressing trials (Table II), with chisel-rock contact detected using a learned dynamical model from the authors' prior work (Leiva et al. 2026) but independently corroborated by human annotations yielding the same success rate. The self-filter uses the robot's kinematic model by design, and the occlusion background model is adapted from external prior work (Yang et al.); these are engineering components, not results derived from the paper's own claims. The RTMDet segmentation model is fine-tuned on a custom dataset from the same testbed, but evaluation is on held-out images and the central pose-generation result does not reduce to segmentation accuracy. Pose-generation thresholds and scoring weights are hand-set and listed in Appendix F; there is no evidence they were fitted to the success-rate outcome. The main caveat is construct validity rather than circularity: the success metric is stable chisel-rock contact during a 5 s simulated press, and the paper explicitly states this 'does not imply that actuating the impact hammer’s end-effector in these situations would result in rock fractures' (Sec. V-C1). Likewise, excluding 6/117 unreachable poses is justified because the control system is independent of the perception pipeline. These are external-validity limitations, not self-referential reductions. The self-citations (Leiva et al. 2026; Cárdenas et al. 2022; Ruiz del Solar et al. 2025) supply components, but the main claims—real-time feasible pose generation at roughly 10 Hz with 72% stable-contact pressing—are independent physical measurements and would stand even if those citations were removed.

Axiom & Free-Parameter Ledger

10 free parameters · 6 axioms · 0 invented entities

The system rests on many hand-set thresholds, scoring weights, and modeling choices, several of which are not numerically reported. No new physical entities are introduced. The central success-rate measurement is an external experimental outcome, so circularity is low, but the many unreported parameters raise the burden on the reader who wants to replicate or falsify the claims.

free parameters (10)
  • theta_soft = 0.5618 rad (32.2 deg)
    Soft deviation angle threshold in Eq. (29); hand-set to discard poses with extreme surface normals.
  • theta_hard = 0.3618 rad (20.7 deg)
    Hard pitch threshold in Eq. (35); hand-set to keep poses within operational strike range.
  • P_min = 10 points
    Minimum cluster size for a candidate pose; hand-set in App. F.
  • scoring_weights_lambda_1_to_4 = 1, 2, 3, 5
    Weights in Eq. (36) for distance to end-effector, temporal consistency, rock volume, and distance to last pose; chosen by hand based on operator feedback, not fitted to the success metric.
  • t_max = 15 (~1.5 s at 10 Hz)
    Cap on temporal consistency score; selected to match pose generation frequency.
  • self_filter_box_sphere_offsets = 0.12 m (Fig. 6); 0.05/0.12 m (Fig. 7)
    Offsets enlarge self-filter masks to cover unmodeled hydraulic hoses; chosen by hand.
  • background_model_thresholds = d_th, v_th, lambda in Eqs. (12)-(14); not numerically reported
    Thresholds for frame-difference and dynamic-matrix update rules; hand-set but values omitted.
  • dbscan_and_voxel_parameters = not reported
    Voxel size, DBSCAN eps, and min_samples for point-cloud rock segmentation are not listed in any table; central to segmentation performance.
  • cross_camera_association_threshold = not reported
    Distance threshold for matching left/right detections of the same rock; necessary for fusion but not specified.
  • grizzly_filter_sphere_radius = not reported
    Radius r in Eq. (22) used for grizzly subtraction; value not given.
axioms (6)
  • domain assumption The scaled Bobcat E10 setup with a 1.5x1.5 m grizzly and ceiling-mounted ZED X cameras is representative of real underground ore-pass impact-hammer environments.
    Section III describes the setup as 'representative of some of the real setups', but no full-scale validation is provided.
  • domain assumption The impact hammer's kinematic model, including the estimated joint offsets, accurately represents the physical machine geometry for self-filtering.
    Section IV-B1 relies on URDF forward kinematics and calibrated offsets, while explicitly acknowledging unmodeled hydraulic hoses and kinematic loops.
  • domain assumption Divergence between a learned dynamical model prediction and the measured configuration reliably indicates chisel-rock contact.
    Section V-C1 uses this learned-model divergence to detect contact; if the model is inaccurate, success and slip rates would be misclassified.
  • domain assumption The depth-based background model of Yang et al. (2004, 2005), with the paper's modifications, correctly handles occlusions caused by the moving hammer.
    Section IV-B2 adapts the pixel/frame-level update rules to depth maps; the modification is not independently validated.
  • domain assumption ZED X stereo depth maps, after bilateral filtering and vendor neural post-processing, are accurate enough for rock segmentation and surface-normal analysis.
    Appendix A lists camera settings; no depth-error characterization is provided.
  • standard math Standard computer vision and geometric algorithms (ICP, DBSCAN, SDF equations, pinhole projection) are correct as used.
    Equations (1), (20), (22), (25) and clustering/ICPs are standard; no formal verification is supplied.

pith-pipeline@v1.3.0-alltime-deepseek · 28566 in / 12734 out tokens · 114938 ms · 2026-08-01T09:29:04.324326+00:00 · methodology

0 comments
read the original abstract

Impact hammers, also known as rock-breakers, are essential machines in mining operations, where they perform secondary reduction. In underground mining, these machines are typically teleoperated, limiting operational efficiency. This paper presents a real-time RGB-D perception pipeline as a step towards automating the operation of hydraulic impact hammers used in mining. The proposed system simultaneously generates operationally feasible rock-breaking poses and a robot-free 3D representation of the workspace. The proposed approach combines image-based instance segmentation with geometric point cloud processing, and operates on embedded hardware at approximately 10 Hz with a total latency of around 675 ms, enabling responsive closed-loop behavior when integrated with a control system. Experimental results in a representative scaled scenario demonstrate that the proposed system is suitable for real-time autonomous impact hammer operation.

Figures

Figures reproduced from arXiv: 2607.20748 by Francisco Leiva, Javier Ruiz-del-Solar, Mart\'in Gallegos, Michelle Cort\'es, Patricio Loncomilla.

Figure 1
Figure 1. Figure 1: The perception pipeline uses RGB-D data and the impact hammer’s [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Real-world experimental setup. mini-excavator to provide a direct view of the overall scene from two different point of views. Both cameras are connected to an NVIDIA Jetson AGX Orin2 via GMSL2 Fakra cables [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Complete perception pipeline diagram, highlighting the primary execution backend utilized for each sub-module. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Calibration layout of the Bobcat E10 mini-excavator. The fixed frames [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Side view of the Bobcat E10 mini-excavator, highlighting simplified [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of the subdivision procedure applied on the meshes of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Self-filter masks obtained via per-link bounding boxes, projected into [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Time evolution of the background models and filled depth images provided by the occlusion management module for the right stereo camera at [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual segmentation process diagram. regions of interest (ROIs) are fixed, camera-dependent and calibrated offline. Moreover, ROIs are rescaled to a resolution of 480×480 pixels before being fed to the segmentation model. Once the segmentation masks are obtained, they are mapped back to the original RGB image resolution (960×600 pixels). Subsequently, their new coordinates are used to extract the segmented… view at source ↗
Figure 10
Figure 10. Figure 10: Example results of the application of point cloud filters. (a) Input [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Point cloud rocks segmentation process. (a) Input point cloud; [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Target rock-breaking poses for point cloud segmented rocks. (a) Raw, [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prioritization of rock-breaking poses based on their distance to the [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Rock-breaking poses prioritized according to the size of the rock to [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: Rock-breaking poses prioritized based on their distance from the last [PITH_FULL_IMAGE:figures/full_fig_p016_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Examples of experimental configurations used for rock segmentation and rock-breaking pose generation. (a) Single rock; (b) two rocks; (c) three [PITH_FULL_IMAGE:figures/full_fig_p017_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Rock-breaking pose generation comparison. Vertically aligned images correspond to the same configuration of rocks in the grizzly. The first row [PITH_FULL_IMAGE:figures/full_fig_p020_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Software architecture diagram. Similarly to Fig. [PITH_FULL_IMAGE:figures/full_fig_p023_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 4 linked inside Pith

  1. [1]

    2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) , volume=

    Real-time multiple objects tracking with occlusion handling in dynamic scenes , author=. 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) , volume=. 2005 , organization=

    2005

  2. [2]

    Proceedings of the ACM 2nd international workshop on Video surveillance & sensor networks , pages=

    Real-time and accurate segmentation of moving objects in dynamic scene , author=. Proceedings of the ACM 2nd international workshop on Video surveillance & sensor networks , pages=

  3. [3]

    Sixth international conference on computer vision (IEEE Cat

    Bilateral filtering for gray and color images , author=. Sixth international conference on computer vision (IEEE Cat. No. 98CH36271) , pages=. 1998 , organization=

    1998

  4. [4]

    arXiv preprint arXiv:1801.09847 , year=

    Open3D: A modern library for 3D data processing , author=. arXiv preprint arXiv:1801.09847 , year=

    Pith/arXiv arXiv

  5. [5]

    James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and George Necula and Adam Paszke and Jake Vander

  6. [6]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

  7. [7]

    arXiv preprint arXiv:2212.07784 , year=

    Rtmdet: An empirical study of designing real-time object detectors , author=. arXiv preprint arXiv:2212.07784 , year=

    Pith/arXiv arXiv

  8. [8]

    IEEE transactions on information theory , volume=

    Least squares quantization in PCM , author=. IEEE transactions on information theory , volume=. 1982 , publisher=

    1982

  9. [9]

    kdd , volume=

    A density-based algorithm for discovering clusters in large spatial databases with noise , author=. kdd , volume=

  10. [10]

    Energies , volume=

    Automatic determination of rock-breaking target poses for impact hammers , author=. Energies , volume=. 2022 , publisher=

    2022

  11. [11]

    ICRA workshop on open source software , volume=

    ROS: an open-source Robot Operating System , author=. ICRA workshop on open source software , volume=. 2009 , organization=

    2009

  12. [12]

    Applied Sciences , volume=

    Haptic teleoperation of impact hammers in underground mining , author=. Applied Sciences , volume=. 2022 , publisher=

    2022

  13. [13]

    Fragblast , volume=

    Automatic detection and breaking system for boulders by use of ccd camera and laser pointer , author=. Fragblast , volume=. 1998 , publisher=

    1998

  14. [14]

    Fragblast , volume=

    Automatic system for boulder breakage using force sensors , author=. Fragblast , volume=. 1999 , publisher=

    1999

  15. [15]

    Autonomous robotic systems , pages=

    Robotics for the mining industry , author=. Autonomous robotic systems , pages=. 2005 , publisher=

    2005

  16. [16]

    IEEE Robotics and Automation Letters , volume=

    Learning to break rocks with deep reinforcement learning , author=. IEEE Robotics and Automation Letters , volume=. 2023 , publisher=

    2023

  17. [17]

    arXiv preprint arXiv:2601.07813 , year=

    Data-driven control of hydraulic impact hammers under strict operational and control constraints , author=. arXiv preprint arXiv:2601.07813 , year=

    arXiv

  18. [18]

    2018 Ninth International Conference on Intelligent Control and Information Processing (ICICIP) , pages=

    Clustering analysis for secondary breaking using a low-cost time-of-flight camera , author=. 2018 Ninth International Conference on Intelligent Control and Information Processing (ICICIP) , pages=. 2018 , organization=

    2018

  19. [19]

    2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) , pages=

    Efficient 3d visual perception for robotic rock breaking , author=. 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) , pages=. 2019 , organization=

    2019

  20. [20]

    Journal of Field Robotics , volume=

    Autonomous robotic rock breaking using a real-time 3D visual perception system , author=. Journal of Field Robotics , volume=. 2021 , publisher=

    2021

  21. [21]

    arXiv preprint arXiv:1804.02767 , year=

    Yolov3: An incremental improvement , author=. arXiv preprint arXiv:1804.02767 , year=

    Pith/arXiv arXiv

  22. [22]

    2021 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM) , pages=

    Robust rock detection and clustering with surface analysis for robotic rock breaking systems , author=. 2021 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM) , pages=. 2021 , organization=

    2021

  23. [23]

    Expert Systems with Applications , volume=

    Detecting rocks in challenging mining environments using convolutional neural networks and ellipses as an alternative to bounding boxes , author=. Expert Systems with Applications , volume=. 2022 , publisher=

    2022

  24. [24]

    Sensor fusion IV: control paradigms and data structures , volume=

    Method for registration of 3-D shapes , author=. Sensor fusion IV: control paradigms and data structures , volume=. 1992 , organization=

    1992

  25. [25]

    arXiv preprint arXiv:2106.13281 , year=

    Brax--a differentiable physics engine for large scale rigid body simulation , author=. arXiv preprint arXiv:2106.13281 , year=

    Pith/arXiv arXiv

  26. [26]

    Journal of Open Source Software , volume=

    PyVista: 3D plotting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK) , author=. Journal of Open Source Software , volume=. 2019 , publisher=

    2019

  27. [27]

    2025 , month=dec, note=

    Ruiz del Solar, Javier and Leiva Castro, Francisco and Romero Vega, Sandra , title=. 2025 , month=dec, note=

    2025