Pith. sign in

REVIEW 2 major objections 7 minor 38 references

Physics-informed Neural Time Fields for Prehensile Object Manipulation

T0 review · 2 major / 7 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 A first-order energy penalty halves training time for physics-informed manipulation planners.

desk verdict Solid incremental PINN manipulation planner; the R^6 pose-space gap and the mislabeled regularizer need fixing before the claims are fully trusted. read the letter →

arxiv 2508.02976 v1 pith:TPDQF2SY submitted 2025-08-05 cs.RO

Hanwen Ren , Ruiqi Ni , Ahmed H. Qureshi This is my paper
classification cs.RO
keywords physics-informedneuralnetworksEikonalequationDirichletenergyprehensilemanipulationtimefieldsre-graspingmotionplanninggeneralizationtonovelobjects
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 establish that physics-informed neural networks can plan prehensile object manipulation—picking up an object and moving it to a target pose in clutter—without expert demonstrations or trial-and-error learning. Its central move is to replace the second-order Laplacian viscosity term used by prior neural time-field planners with a first-order Dirichlet energy penalty, cutting training time roughly in half (about 1.3 hours versus 2.6–3.0 hours) while keeping planning quality essentially unchanged. A second contribution is a shape encoder that lets the learned time field generalize to objects never seen in training, and a recursive regrasping procedure that lets the robot change grasps mid-task. The result, if correct, is a practical planner that produces collision-free, kinematically feasible manipulation trajectories in about 0.05–0.07 seconds.

What carries the argument

The load-bearing object is the Dirichlet energy $E_D(T)=\int_\Omega |\nabla T|^2\,d\Omega$, a first-order penalty whose minimization is equivalent to solving the Laplace equation $\Delta T=0$; it replaces the second-order Laplacian viscosity term of P-NTFields, giving the same smoothing effect at half the derivative cost. Around it, the architecture pairs a Fourier-feature pose encoder with a symmetric max/min operator (enforcing $T(p_s,p_g)=T(p_g,p_s)$) and a PointNet++ shape encoder that maps an object's 64-point point cloud to a latent code, which is what allows the field to generalize to unseen objects. The training loss combines an isotropic, ratio-symmetric speed-matching term $S^*/S + S/S^* - 2$ at start and goal with the Dirichlet penalty under progressive speed scheduling. Trajectories are extracted by bidirectional gradient descent in pose space, and the OManip wrapper searches over Contact-GraspNet grasps, checks inverse-kinematics feasibility, and recursively inserts a decoupled intermediate pose when a single grasp fails—this recursion is what produces re-grasping behavior.

What would settle it

Take a long, asymmetric object and set start and goal orientations differing by a large rotation (for example, 180° about the vertical axis) inside the U-shaped tunnel, where the object must rotate within confined space. If the planned trajectory, when executed as a swept point cloud, collides with the tunnel walls even though the raw pose-space path is collision-free, the Euclidean pose-space assumption is refuted. A more direct check is to compare the trajectory's Euclidean length against its length computed with a proper SE(3) geodesic metric; a systematic gap would show that the gradients do not track minimal valid rigid-body motion.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the Eikonal equation $\|\nabla_{p_g} T(p_s, p_g)\| = 1/S(p_g)$ can be trained for object manipulation with a cost function that augments the speed-matching loss with the Dirichlet energy $\int_\Omega |\nabla T|^2\,d\Omega$ instead of a Laplacian viscosity term, and that this substitution is sufficient to obtain stable training, fast convergence, and smooth trajectories. Because the Dirichlet energy requires only first derivatives of the network, the method avoids the computational overhead of the second-order viscosity term and trains in about 1.3 hours per environment—roughly half the time of the P-NTFields baseline—while matching its success rates (98.8% versus 98.4% in the tabletop environment, for example) and even exceeding its planning speed. The paper further claims that feeding a PointNet++-encoded point cloud of the object into the time-field network makes the planner generalize to five held-out YCB objects without retraining, and that a recursive search over grasps with intermediate poses produces multimodal plans in which the robot re-grasps the object when a single grasp cannot realize the trajectory. The method is demonstrated in three simulated environments and on a real robot in a cabinet.

Load-bearing premise

The planner treats an object's six-dimensional pose as a Euclidean point and performs ordinary gradient descent in raw pose coordinates, without specifying how orientation is parametrized or how the gradient step respects the rotation group $\mathrm{SO}(3)$; if straight-line steps in this $\mathbb{R}^6$ space do not correspond to valid object rotations, the resulting trajectories are not truly kinematically feasible.

Editorial extensions

If this is right

  • Training a physics-informed manipulation planner drops from roughly 2.6–3.0 hours to about 1.3 hours, making per-environment retraining practical for deployment.
  • Planning queries complete in about 0.05–0.07 seconds, substantially faster than the RRT-Connect and JIST baselines in these environments.
  • The learned time field transfers to five unseen YCB objects without fine-tuning, so the planner does not need per-object retraining.
  • When a single grasp cannot realize the full trajectory, the recursive intermediate-pose search yields multi-step plans with re-grasping, extending PINN planning to tasks that require changing contacts.
  • Because the Dirichlet penalty is a drop-in replacement for the viscosity term, the same substitution applies to the broader NTFields/P-NTFields family of physics-informed planners, not just the manipulation setting presented here.

Reading between the lines

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

  • The paper's Euclidean pose-space metric is the main unexamined assumption; a natural extension is to endow the pose encoder with an SE(3)-aware distance, such as a geodesic or quaternion-based orientation metric, and test whether planning quality and trajectory length improve on large-rotation tasks.
  • The speed function in Eq. 5 is hand-defined from point-cloud distance and a reachability indicator; a testable extension is to learn the speed function jointly with the time field, which could remove the clipping parameters and reduce manual tuning.
  • The regrasping search enumerates grasps in stability order and decouples a single rotation dimension at a time; one could instead have the network predict the intermediate pose directly, which might shorten the recursive search.
  • If the reported planning times hold across more varied environments, PINN-based planning becomes a credible real-time alternative for tasks where demonstrations are unavailable, including the non-prehensile pushing and tool-use tasks the authors list as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The paper introduces POM-NeTF, a physics-informed neural network that solves the Eikonal equation in object-pose space for prehensile object manipulation. It replaces the Laplacian viscosity regularization of P-NTFields with a first-order gradient penalty called Dirichlet energy minimization, uses a PointNet++ encoder to generalize across YCB objects, and plans by bidirectional gradient descent over poses, with recursive re-grasping when inverse kinematics fails. Experiments in three simulated environments and a real cabinet report planning times around 0.05-0.07 s, training times around 1.3 h versus 2.6-3.0 h for P-NTFields, high success rates, and generalization to unseen objects.

Significance. If the claims hold, POM-NeTF is a meaningful advance: it is the first demonstration that PINN-based planning can handle prehensile manipulation without expert demonstrations, with a large training-time reduction and near-real-time planning. The empirical work is substantial: 1500 simulated test cases, three environments, real-robot comparisons, and explicit data-generation and training times. The comparisons against sampling-based baselines are framed fairly in that all methods use the same OManip post-processing. The main contributions, faster training via first-order regularization and architecture-level object generalization, are clearly stated and independently evaluated against baselines. However, the validity of the central claims depends on two technical points that are not yet established: the relation between the implemented loss and the Dirichlet energy of Eq. (3), and the metric and parametrization of the six-dimensional pose space. These issues are local and fixable, but they are load-bearing for the stated claims of faster convergence and kinematically feasible trajectories.

major comments (2)
  1. [§IV-A, Eq. (3), Eq. (7)] The paper claims that replacing the Laplacian viscosity term with Dirichlet energy minimization halves training time, but Eq. (7) does not implement the Dirichlet energy of Eq. (3). Equation (3) is the integral of |∇T|² over a domain, whereas Eq. (7) penalizes only |∇_{p_g} T(p_s,p_g)|², a partial derivative with respect to the goal pose, and the manuscript never defines the domain Ω or the full gradient over the two-pose input space. The stated equivalence between minimizing E_D(T) and solving the Laplace equation is therefore not applicable to the actual loss. If the intention is to use only the goal-pose gradient as a cheap viscosity surrogate, that needs to be stated and mathematically justified; otherwise the training-time comparison is explained by a different mechanism than the one claimed.
  2. [§III-A, §IV-E, Eq. (8)] The pose space is described as p ∈ R^6, but the paper never specifies how orientation is represented or which metric is used on SO(3). Equation (8) performs gradient descent directly in these raw coordinates and terminates using the Euclidean distance ||p_s' − p_g'||_2, while Eq. (5) computes distances between point clouds in Euclidean workspace. If orientation is encoded with Euler angles or axis-angle, the flat Euclidean metric does not coincide with a natural Riemannian metric on SO(3), the chart has singularities, and Euclidean distance in pose space does not measure physical rotation. Consequently the learned time field is not the solution of an Eikonal equation on SE(3), its characteristics may not correspond to valid object rotations, and the reported 'Length (m)' numbers in Tables I and III do not have a clear physical meaning. The authors should specify the orientation parametrization and the metric, and validate that the planned pose curves are continuous and physically realizable in SE(3), for example by comparing the rotational part of planned paths with the SO(3) geodesic distance and by evaluating behavior near chart singularities. Without this, the 'kinematically feasible' claim and the generalization results are not fully supported.
minor comments (7)
  1. [Algorithm 1, lines 5-12] The indentation of the pseudocode is ambiguous: line 12 `return RMP(σr)` appears to be placed inside the failure branch, where σr may not be defined if the inverse kinematics mapping failed.
  2. [Table III] The column header 'SR (%) ↓' should use an upward arrow, since higher success rates are better.
  3. [Fig. 1 caption] The caption says '(Left) the tabletop ... (Middle) ... and (Left) the cabinet'; the last parenthetical should read '(Right)'.
  4. [§III-A] The heading contains a typo: 'Problem Defination' should be 'Problem Definition'.
  5. [§II] The related-work section contains the typo 'through trail and error' and should read 'trial and error'.
  6. [Table II and §V-A] The training times in Table II are single numbers with no variance or number of seeds; reporting multiple training runs would strengthen the claim that Dirichlet energy regularization halves training time.
  7. [§IV-F, Table III] The real-robot experiments in Table III report means and standard deviations but do not state the number of trials per object or per start-goal pair; this information is needed to interpret the success rates and variance.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; central claims are empirically evaluated against external baselines and real-world experiments.

full rationale

The paper builds on the authors' prior NTFields and P-NTFields work, citing those papers for the Eikonal formulation, the isotropic speed-ratio loss (Eq. 6), the progressive speed scheduling, and the symmetric operator. These are adopted as components and baselines, not as load-bearing justifications that force the central conclusions. The key novel steps — Dirichlet energy regularization (Eq. 3/7), the object-shape encoder, the improved speed model with IK reachability (Eq. 5), and the recursive OManip planning algorithm — are clearly specified and independently testable. The training-time comparison against P-NTFields is a controlled empirical experiment in which the only substantive difference is the Laplacian viscosity versus the first-order Dirichlet energy term; the quoted equivalence of Dirichlet-energy minimization to the Laplace equation is a standard mathematical fact, not a self-referential premise. The planning metrics (time, length, success rate) are computed from actual test executions, including 500 test cases per environment and real-robot trials, and are compared against external sampling-based baselines (RRT-Connect, JIST). No test-time constant or parameter is fitted to the reported success metrics, and no prediction reduces by construction to an input of the method. The paper's treatment of object pose as Euclidean R^6 in Eq. 8 is a potential geometrical correctness concern, but it is not a circularity: the ground-truth speed is computed from geometry and IK, not from the network's own outputs, and the planning pipeline is empirically validated end-to-end. Overall, the derivation chain is self-contained; the self-citations are ancillary and do not make the results equivalent to their inputs.

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

The central claim rests on the Eikonal/reachability modeling and on hyperparameters carried over from prior work, not on any new free parameters fit to the quoted success metrics. The main uncharged assumption is the Euclidean treatment of object pose space, which is the weakest unexamined premise.

free parameters (6)
  • s_const = not reported
    Speed scaling constant in Eq.5; chosen by hand, likely from prior work.
  • d_min and d_max = not reported
    Clipping limits in the speed function Eq.5; hand-chosen to keep speed between bounds.
  • epsilon (gradient penalty weight) = not reported
    Weight on the pointwise |∇pg T|^2 term in Eq.7; controls regularization strength.
  • eta (trajectory step size) = not reported
    Step size in the bidirectional trajectory updates of Eq.8.
  • d_s (convergence threshold) = not reported
    Threshold on ||p_s' - p_g'|| that stops trajectory generation.
  • Progressive speed scheduling parameters = not reported
    Adopted verbatim from P-NTFields; not specified in this paper.
assumptions (5)
  • domain assumption The Eikonal equation with speed function S is a valid model for object manipulation planning.
    Introduced in Section III-B and used throughout; assumes the shortest-path wavefront model transfers from robot configuration space to object pose space.
  • standard math Minimizing Dirichlet energy (Eq.3) is equivalent to solving the Laplace equation and provides regularization similar to the Laplacian viscosity term.
    Stated in Section IV-A; however, the implemented loss in Eq.7 uses only a pointwise |∇pg T|^2, not the full integral energy, so this axiom is invoked but not realized in the implementation.
  • ad hoc to paper Object pose space can be treated as Euclidean R6 with L2 distance and raw coordinate gradient descent.
    Used in Eq.8 and in the ground-truth speed d(Xpo, Xobs); no justification is given for the SO(3) rotational component.
  • domain assumption The ground-truth speed function Eq.5 with the IK reachability indicator provides a faithful model for kinematically feasible manipulation.
    Defined in Section IV-C; the binary indicator IKr makes speed discontinuous and is a heuristic for reachability.
  • domain assumption Progressive speed scheduling from P-NTFields applies unchanged to the pose-space setting.
    Adopted in Section IV-D without description; relies on the authors' prior self-cited work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed Neural Time Fields for Prehensile Object Manipulation." pith.science (2026). https://pith.science/paper/TPDQF2SY

@misc{pith2026250802976,
  author       = {Pith},
  title        = {Pith review of: Physics-informed Neural Time Fields for Prehensile Object Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TPDQF2SY}},
  note         = {Machine review of arXiv:2508.02976}
}
read the original abstract

Object manipulation skills are necessary for robots operating in various daily-life scenarios, ranging from warehouses to hospitals. They allow the robots to manipulate the given object to their desired arrangement in the cluttered environment. The existing approaches to solving object manipulations are either inefficient sampling based techniques, require expert demonstrations, or learn by trial and error, making them less ideal for practical scenarios. In this paper, we propose a novel, multimodal physics-informed neural network (PINN) for solving object manipulation tasks. Our approach efficiently learns to solve the Eikonal equation without expert data and finds object manipulation trajectories fast in complex, cluttered environments. Our method is multimodal as it also reactively replans the robot's grasps during manipulation to achieve the desired object poses. We demonstrate our approach in both simulation and real-world scenarios and compare it against state-of-the-art baseline methods. The results indicate that our approach is effective across various objects, has efficient training compared to previous learning-based methods, and demonstrates high performance in planning time, trajectory length, and success rates. Our demonstration videos can be found at https://youtu.be/FaQLkTV9knI.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 29 canonical work pages

  1. [1]

    Trends and challenges in robot manipulation

    Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446):eaat8414, 2019

    work page 2019

  2. [2]

    Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics , 378:686–707, 2019

    2019

  3. [3]

    Ntfields: Neural time fields for physics-informed robot motion planning

    Ruiqi Ni and Ahmed H Qureshi. Ntfields: Neural time fields for physics-informed robot motion planning. arXiv preprint arXiv:2210.00120, 2022

    arXiv 2022

  4. [4]

    Progressive learning for physics- informed neural motion planning

    Ruiqi Ni and Ahmed H Qureshi. Progressive learning for physics- informed neural motion planning. arXiv preprint arXiv:2306.00616 , 2023

    arXiv 2023

  5. [5]

    Constrained sampling- based planning for grasping and manipulation

    Jinwook Huh, Bhoram Lee, and Daniel D Lee. Constrained sampling- based planning for grasping and manipulation. In 2018 IEEE International Conference on Robotics and Automation (ICRA) , pages 223–230. IEEE, 2018

    work page 2018

  6. [6]

    Robot kinematics: Forward and inverse kinematics

    Serdar Kucuk and Zafer Bingul. Robot kinematics: Forward and inverse kinematics. INTECH Open Access Publisher London, UK, 2006

    work page 2006

  7. [7]

    Search- based planning for manipulation with motion primitives

    Benjamin J Cohen, Sachin Chitta, and Maxim Likhachev. Search- based planning for manipulation with motion primitives. In 2010 IEEE international conference on robotics and automation , pages 2902–2908. IEEE, 2010

    work page 2010

  8. [8]

    Manipulation with Shared Grasping

    Yifan Hou, Zhenzhong Jia, and Matthew T Mason. Manipulation with shared grasping. arXiv preprint arXiv:2006.02996 , 2020

    work page Pith review arXiv 2006

Show all 38 references
  1. [9]

    Visual detection of opportunities to exploit contact in grasping using contextual multi-armed bandits

    Clemens Eppner and Oliver Brock. Visual detection of opportunities to exploit contact in grasping using contextual multi-armed bandits. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 273–278. IEEE, 2017

    2017

  2. [10]

    Pick and place without geometric object models

    Marcus Gualtieri, Andreas Ten Pas, and Robert Platt. Pick and place without geometric object models. In 2018 IEEE international conference on robotics and automation (ICRA) , pages 7433–7440. IEEE, 2018

    2018

  3. [11]

    Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations

    Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters , 5(3):4978–4985, 2020

    2020

  4. [12]

    A framework for behavioural cloning

    Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence 15 , pages 103–129, 1995

    1995

  5. [13]

    Algorithms for inverse reinforce- ment learning

    Andrew Y Ng, Stuart Russell, et al. Algorithms for inverse reinforce- ment learning. In Icml, volume 1, page 2, 2000

    2000

  6. [14]

    Learn- ing manipulation actions from human demonstrations

    Tim Welschehold, Christian Dornhege, and Wolfram Burgard. Learn- ing manipulation actions from human demonstrations. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3772–3777. IEEE, 2016

    2016

  7. [15]

    Neural descriptor fields: Se (3)-equivariant object representations for manipulation

    Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenen- baum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (IC...

    2022

  8. [16]

    Useek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation

    Zhengrong Xue, Zhecheng Yuan, Jiashun Wang, Xueqian Wang, Yang Gao, and Huazhe Xu. Useek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 1715–1722. IEEE, 2023

    2023

  9. [17]

    Qt-opt: Scalable deep reinforce- ment learning for vision-based robotic manipulation

    Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakr- ishnan, Vincent Vanhoucke, et al. Qt-opt: Scalable deep reinforce- ment learning for vision-based robotic manipulation. arXiv preprint arXiv:18...

    2018 arXiv

  10. [18]

    Hacman: Learning hybrid actor-critic maps for 6d non- prehensile manipulation

    Wenxuan Zhou, Bowen Jiang, Fan Yang, Chris Paxton, and David Held. Hacman: Learning hybrid actor-critic maps for 6d non- prehensile manipulation. In Conference on Robot Learning , pages 241–265. PMLR, 2023

    2023

  11. [19]

    Synergistic task and motion planning with reinforcement learning-based non-prehensile actions

    Gaoyuan Liu, Joris De Winter, Denis Steckelmacher, Roshan Kumar Hota, Ann Nowe, and Bram Vanderborght. Synergistic task and motion planning with reinforcement learning-based non-prehensile actions. IEEE Robotics and Automation Letters , 8(5):2764–2771, 2023

    2023

  12. [20]

    Rearrangement with nonprehensile manipulation using deep reinforcement learning

    Weihao Yuan, Johannes A Stork, Danica Kragic, Michael Y Wang, and Kaiyu Hang. Rearrangement with nonprehensile manipulation using deep reinforcement learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA) , pages 270–277. IEEE, 2018

    2018

  13. [21]

    Multi-stage rein- forcement learning for non-prehensile manipulation

    Dexin Wang, Faliang Chang, and Chunsheng Liu. Multi-stage rein- forcement learning for non-prehensile manipulation. arXiv preprint arXiv:2307.12074, 2023

    2023 arXiv

  14. [22]

    Beyond pick-and-place: Tackling robotic stacking of diverse shapes

    Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byra- van, Abbas Abdolmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In Conference on Robot...

    2022

  15. [23]

    Deep reinforcement learning for robotic manipulation

    Shixiang Gu, Ethan Holly, Timothy P Lillicrap, and Sergey Levine. Deep reinforcement learning for robotic manipulation. arXiv preprint arXiv:1610.00633, 1:1, 2016

    2016 arXiv

  16. [24]

    Self-organizing neural networks integrating domain knowledge and reinforcement learning

    Teck-Hou Teng, Ah-Hwee Tan, and Jacek M Zurada. Self-organizing neural networks integrating domain knowledge and reinforcement learning. IEEE transactions on neural networks and learning systems, 26(5):889–902, 2014

    2014

  17. [25]

    A fast marching level set method for monotonically advancing fronts

    James A Sethian. A fast marching level set method for monotonically advancing fronts. proceedings of the National Academy of Sciences , 93(4):1591–1595, 1996

    1996

  18. [26]

    A fast marching algorithm for the factored eikonal equation

    Eran Treister and Eldad Haber. A fast marching algorithm for the factored eikonal equation. Journal of Computational physics , 324:210–225, 2016

    2016

  19. [27]

    Random features for large-scale kernel machines

    Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007

    2007

  20. [28]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich- Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Pro...

    2020

  21. [29]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems , 30, 2017

    2017

  22. [30]

    Contact-graspnet: Efficient 6-dof grasp generation in cluttered scenes

    Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, and Di- eter Fox. Contact-graspnet: Efficient 6-dof grasp generation in cluttered scenes. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages 13438–13444. IEEE, 2021

    2021

  23. [31]

    Riemannian motion policies

    Nathan D Ratliff, Jan Issac, Daniel Kappler, Stan Birchfield, and Dieter Fox. Riemannian motion policies. arXiv preprint arXiv:1801.02854, 2018

    2018 arXiv

  24. [32]

    Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols

    Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols. arXiv preprint arXiv:1502.03143 , 2015

    2015 arXiv

  25. [33]

    Fast marching farthest point sampling

    Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003

    2003

  26. [34]

    Rrt-connect: An efficient approach to single-query path planning

    James J Kuffner and Steven M LaValle. Rrt-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065) , vol- ume 2, pages 995–...

    2000

  27. [35]

    Fast, anytime motion planning for prehensile manipulation in clutter

    Andrew Kimmel, Rahul Shome, Zakary Littlefield, and Kostas Bekris. Fast, anytime motion planning for prehensile manipulation in clutter. In 2018 IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids), pages 1–9. IEEE, 2018

    2018

  28. [36]

    The open motion planning library

    Ioan A Sucan, Mark Moll, and Lydia E Kavraki. The open motion planning library. IEEE Robotics & Automation Magazine , 19(4):72– 82, 2012

    2012

  29. [37]

    Fcl: A general purpose library for collision and proximity queries

    Jia Pan, Sachin Chitta, and Dinesh Manocha. Fcl: A general purpose library for collision and proximity queries. In 2012 IEEE International Conference on Robotics and Automation , pages 3859–

    2012

  30. [38]

    Sampling-based algorithms for optimal motion planning

    Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846–894, 2011

    2011

Pith tools

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