Pith. sign in

REVIEW 5 major objections 6 minor 41 references

PRISM: Pointcloud Reintegrated Inference via Segmentation and Cross-attention for Manipulation

T0 review · 5 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 PRISM claims that segmenting a point cloud into object clusters and letting joint states attend over them lifts from-scratch imitation learning to an 82.0% average success rate across six RoboTwin tasks, above DP3's 58.4%.

desk verdict PRISM is a plausible engineering win with a load-bearing but underspecified segmentation step; send to review, insist on reproducible evaluation. read the letter →

arxiv 2507.04633 v1 pith:YPMYQSNM submitted 2025-07-07 cs.RO

Daqi Huang , Zhehao Cai , Yuzhi Hao , Zechen Li , Chee-Meng Chew This is my paper
classification cs.RO
keywords imitationlearningrobotmanipulationpointcloudsegmentationDBSCANcross-attentiondiffusionpolicymulti-stagesensorfusion
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

PRISM is an end-to-end imitation-learning policy that works directly from raw point clouds and robot joint states, without pretrained encoders or external datasets. The paper's claim is that by first cutting the point cloud into object-sized clusters and then letting each arm's joint state attend over those clusters, the policy concentrates on the right object at each stage of a task. On 100 demonstrations per task, PRISM reports an 82.0% average success rate across six dual-arm RoboTwin tasks, compared with 58.4% for the 3D Diffusion Policy and 22.3% for the 2D Diffusion Policy. A sympathetic reader would care because the result suggests that simple unsupervised segmentation plus cross-modal attention, rather than larger pretrained representations, may be what makes imitation policies reliable in cluttered and multi-stage manipulation.

What carries the argument

The load-bearing object is the segmentation-embedding unit that converts an unordered scene point cloud into a small set of object-level tokens. DBSCAN (density-based spatial clustering with noise) partitions the cloud by geometric proximity into object-centric clusters; a shared-parameter PointNet with LayerNorm and max-pooling maps each cluster to a local descriptor; a transformer encoder then reweights these descriptors through self-attention to capture inter-object relations; and the resulting tokens are queried by per-arm joint-state embeddings in a cross-attention layer that outputs the conditioning signal for a DDIM-sampled diffusion action decoder. This tokenization is what lets the policy attend to the cup at one phase and the mat at the next, the behavior shown in the attention heat maps.

What would settle it

Compare PRISM's success rate on a task where two objects touch or overlap so that DBSCAN merges them into a single cluster, or where one object is fragmented by the noise threshold; if the reported 82.0% average depends on clean clusterings, the score should collapse or swing widely. A cheaper check is to sweep the unstated DBSCAN epsilon and minPts over a sensible range on Pick Apple Messy and see whether the 71.8% result is a plateau or a knife-edge.

Watch

Extended reading notes

Core claim

The central discovery is that object-centric tokenization changes how well a diffusion policy can imitate manipulation. PRISM runs DBSCAN on the raw point cloud to produce K clusters, embeds each cluster with a shared-parameter PointNet that uses LayerNorm instead of batchnorm, reshapes the cluster features with a transformer encoder, and then fuses the result with the robot's joint states through decoupled cross-attention in which left and right arm states carry positional embeddings. The fused features condition a diffusion decoder that generates smooth action sequences. In the paper's experiments the full system averages 82.0% success over six tasks, the transformer encoder contributes a 6.5-point gain in ablation (75.5% without it), and the gains are largest in the cluttered Pick Apple Messy task, where PRISM reaches 71.8% versus 16.4% for DP3.

Load-bearing premise

The method stands on DBSCAN reliably splitting each scene into one cluster per object using a single fixed set of clustering parameters, but the paper reports no values for those parameters and no analysis of clustering quality, even though it asserts the same settings work across all six tasks.

Editorial extensions

If this is right

  • An imitation policy for dual-arm manipulation can be trained end-to-end from raw point clouds and joint states in about 100 demonstrations per task, with no pretrained vision backbone or external dataset.
  • Segmenting the scene into object tokens before feature extraction removes the distraction that holistic point-cloud encoders suffer from in clutter; the grouped cluttered-task average is 66.1% versus 28.6% for DP3.
  • The transformer encoder's inter-object self-attention is worth 6.5 points of average success (82.0% versus 75.5%), so relational context between objects is a measurable part of the gain.
  • Cross-attention with per-arm positional embeddings lets the model shift its focus between sub-goals, which the paper demonstrates on Blocks Stack and Empty Cup Place where attention moves from the first object to the second mid-episode.
  • On the six-task average the recipe of segmentation plus cross-attention plus diffusion outperforms both the 3D baseline (58.4%) and the 2D image baseline (22.3%), indicating the mechanism, not additional data, drives the improvement.

Reading between the lines

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

  • Editorial extension: because DBSCAN carries no semantics, the same tokenization should transfer to novel objects and layouts without retraining; a direct test would be measuring success as the set of distractor objects is changed between evaluation and training, which the paper does not report.
  • Editorial extension: the cross-attention design can be reused as the grounding head of a language-conditioned policy, using a language embedding as the query in place of joint states — the direction the authors gesture at in their conclusion but do not implement.
  • Editorial extension: the paper's own low-data discussion implies a lighter relational module than a full transformer encoder could capture the same inter-object gain; comparing a single self-attention layer against the full encoder at 20 and 50 demonstrations would settle the trade-off the authors observe but leave open.
  • Editorial inference: if the DBSCAN parameters really are task-invariant as claimed, cluster stability under camera viewpoint changes should hold; a stress test with overlapping objects that merge into one cluster would reveal whether the 82.0% average depends on favorable segmentation luck.
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

5 major / 6 minor

Summary. The paper proposes PRISM, an end-to-end imitation learning policy that takes raw point clouds and robot joint states as input. It first segments the point cloud into object-centric clusters using DBSCAN, encodes each cluster with a shared PointNet, refines the object tokens with a transformer encoder, fuses them with proprioceptive state via per-arm cross-attention, and generates action sequences with a diffusion decoder. The method is trained from scratch on 100 demonstrations per task and evaluated on six RoboTwin tasks. The reported overall average success rate is 82.0%, compared with 58.4% for DP3 and 22.3% for Diffusion Policy, and an ablation attributes 6.5 percentage points of the gain to the transformer encoder. The central claim is that unsupervised density-based segmentation combined with cross-attention yields effective selective focus in cluttered and multi-stage manipulation scenarios.

Significance. If the reported results hold, PRISM would be a strong and conceptually simple contribution: it demonstrates that a purely geometric, unsupervised clustering front-end plus cross-attention can rival or exceed specialized 3D visuomotor policies without pretrained backbones or external datasets. The paper is clearly written in its architecture description, and the authors provide code and attention heatmaps that make the proposed mechanism interpretable. The main significance, however, depends on the credibility of the empirical evaluation, which currently suffers from a checkpoint-selection protocol that inflates performance and from missing hyperparameter and baseline details. These issues are fixable but must be addressed before the comparison can be taken as evidence of the method's advantage.

major comments (5)
  1. [IV-A, Evaluation metrics] The reported success rates are obtained by selecting, for each seed, the checkpoint with the highest success rate out of 20 checkpoints saved every 200 epochs, and then averaging these peak rates across five seeds. This is a form of test-set peeking: it reports the maximum over training time of a noisy evaluation, which can substantially overstate true policy performance and understate variance. The authors should either report results from a fixed training schedule (e.g., the final checkpoint) or use a validation-based early-stopping rule, and they should specify which protocol was applied to the DP3 and DP baselines. Without this, the headline comparison in Table III is not a fair or reliable estimate of expected performance.
  2. [III-A, Eq. (1)] The segmentation module is load-bearing for the paper's explanation of the gains, but the manuscript never reports the DBSCAN parameters eps, minPts, or the noise threshold λ_noise, despite asserting that a unified setting works across all tasks. Moreover, Eq. (1) is the HDBSCAN stability criterion (summing λ_core - λ_noise over points in a cluster), not the standard DBSCAN formulation described in the text, so the actual clustering algorithm is ambiguous. The authors should report the exact hyperparameters, clarify whether DBSCAN or HDBSCAN is used, and provide clustering-quality evidence (e.g., agreement with ground-truth object instances or a sensitivity study), especially for tasks like ECPM and PAM where objects can touch or overlap and merged clusters would corrupt the object embeddings.
  3. [IV-A, Baselines; Table III] The evaluation protocol for the baselines is not reported. It is unclear whether DP3 and DP used the same camera view, point-cloud sampling density, number of demonstrations, evaluation rollouts, and checkpoint-selection rule as PRISM. Since the central claim is an empirical performance comparison, the authors must specify the baseline training/evaluation configuration or cite the exact setup from the original papers. Without this information, the 82.0% vs. 58.4% vs. 22.3% comparison cannot be independently assessed or reproduced.
  4. [Abstract; Fig. 1 caption; Section IV] The abstract and Fig. 1 caption claim effectiveness in "diverse simulation and real-world tasks," but Section IV describes only simulated RoboTwin experiments; no real-world results are presented anywhere in the manuscript. If real-world experiments were conducted, they should be included or referenced; otherwise the claim should be removed or qualified to avoid overstating the empirical scope.
  5. [III-D; IV-A] Several hyperparameters essential for reproducibility are omitted: the action horizon n, the number of diffusion denoising steps, the DDIM sampling steps, the number of training epochs, learning rate, batch size, and the point-cloud preprocessing (e.g., voxelization or random downsampling). These details are needed to reproduce the reported numbers and to assess whether the diffusion module's behavior is standard. The authors should include a hyperparameter table or state the missing values in the text.
minor comments (6)
  1. [III-B heading] The heading "Transfomer Encoder" contains a typo and should read "Transformer Encoder."
  2. [I, Introduction] Several typos and grammatical errors appear in the introduction (e.g., "increasely," "cruicial," "vison"); a careful proofreading pass is recommended.
  3. [IV-D, Table II] The ablation table reports only point estimates without standard deviations or per-seed results, so the 6.5-point improvement from the transformer encoder is not statistically supported; adding error bars or per-seed data would strengthen the claim.
  4. [III-A, Eq. (1) and surrounding text] The notation in the cluster definition is inconsistent: C_t^k is defined with index i in {1,...,N_k} but written as {p_j^t}, and the superscript/subscript roles of k and t are used loosely; please unify the notation.
  5. [IV-B, Fig. 4] The attention heatmaps are qualitative; a quantitative analysis (e.g., attention weight on the target object versus distractors averaged over episodes) would more convincingly demonstrate selective focus.
  6. [III-D, Eq. (6)] The scheduler coefficients β, δ, and σ in Eq. (6) are not defined, and the relationship to the DDIM scheduler is not explained; please clarify the denoising update rule.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: PRISM's central claims are empirical performance comparisons built from standard components with external citations, not derivations that reduce to their own inputs.

full rationale

PRISM makes no formal derivation whose conclusion is equivalent to its premises; the load-bearing claims are measured success rates on RoboTwin tasks. The segmentation embedding (Section III-A), transformer encoder (Section III-B), cross-attention (Section III-C), and diffusion decoder (Section III-D) are standard components with external citations (DBSCAN [28]–[30], PointNet [25], attention [32], DDPM/DDIM [35], [37]); none of these citations is authored by the present paper, so no self-citation chain is present. The reported 82.0% overall average is a benchmark result compared with DP3 and DP under the same protocol, not a quantity derived from a fitted parameter or from a self-referential uniqueness claim. The only protocol concern is the selection of the per-seed best checkpoint for reporting ('the checkpoint yielding the highest success rate is identified'), which is an evaluation and statistical bias rather than a case where a prediction reduces by construction to an input; it does not make the central performance claim definitionally identical to the training data. The DBSCAN/HDBSCAN inconsistency in Eq. (1) and the unreported clustering hyperparameters are correctness and reproducibility risks, not circular steps. Accordingly, no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method's novel machinery rests on an unverified clustering premise and an unproven attention-focusing premise, plus an evaluation protocol (best-checkpoint-per-seed) that can inflate success rates. No new physical or mathematical entities are postulated; all building blocks are standard. The lack of reported hyperparameter values makes the ledger incomplete from the paper text alone.

free parameters (4)
  • DBSCAN eps (neighborhood radius) = Not reported
    Hand-chosen clustering radius; segmentation quality and thus the object-tokenization of the whole method depends on it. The paper does not report the value despite claiming a unified setting across tasks. (Section III-A)
  • DBSCAN minPts = Not reported
    Hand-chosen minimum points threshold; controls which points are core, border, or noise, and therefore the stability and count of clusters. (Section III-A)
  • Noise threshold λ_noise in Eq. (1) = Not reported
    Introduced in the stability criterion but not defined or used in the implementation; appears to be an ad hoc parameter carried over from HDBSCAN-style notation. (Section III-A, Eq. 1)
  • Diffusion action horizon and denoising steps = Not reported
    The action sequence horizon and number of diffusion steps are not stated; these affect action smoothness and inference speed, which the paper claims as an advantage. (Section III-D)
assumptions (5)
  • domain assumption DBSCAN density clustering will decompose the raw point cloud into object-centric clusters without per-task tuning of eps or minPts.
    Section III-A: the raw point cloud is divided into K object-centric clusters using DBSCAN; the paper claims a unified noise threshold and minPts work for all tasks but never reports the values. Objects that touch or overlap may be merged or split, corrupting the object-token premise.
  • ad hoc to paper Cross-attention between joint states and object tokens will cause the policy to focus on task-relevant objects.
    Section III-C: the paper asserts the mechanism enables dynamic focus; this is a modeling hypothesis with no proof. The attention heat maps in Fig. 4 are post-hoc visualizations, not quantitative evidence that attention causes the success.
  • domain assumption The RoboTwin simulation benchmark is representative of real cluttered and multi-stage manipulation; real-world validation is not provided.
    Section IV-A describes the simulation setup; the conclusion says results are 'simulation-validated' and 'ready for future hardware iterations'. The Fig. 1 caption's 'real-world tasks' claim is unsupported.
  • domain assumption Selecting the best checkpoint per seed is a valid estimator of policy performance and was applied identically to baselines.
    Section IV-A: 'the checkpoint yielding the highest success rate is identified' for each seed. If the baselines did not use the same protocol, the comparison is unfair; even if they did, the reported success rates are optimistic point estimates.
  • standard math Diffusion model with MSE noise-prediction loss and DDIM sampling produces smooth, executable actions.
    Section III-D: standard denoising diffusion formulation as in DDPM/DDIM [35,37]. This is a well-established background result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PRISM: Pointcloud Reintegrated Inference via Segmentation and Cross-attention for Manipulation." pith.science (2026). https://pith.science/paper/YPMYQSNM

@misc{pith2026250704633,
  author       = {Pith},
  title        = {Pith review of: PRISM: Pointcloud Reintegrated Inference via Segmentation and Cross-attention for Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YPMYQSNM}},
  note         = {Machine review of arXiv:2507.04633}
}
read the original abstract

Robust imitation learning for robot manipulation requires comprehensive 3D perception, yet many existing methods struggle in cluttered environments. Fixed camera view approaches are vulnerable to perspective changes, and 3D point cloud techniques often limit themselves to keyframes predictions, reducing their efficacy in dynamic, contact-intensive tasks. To address these challenges, we propose PRISM, designed as an end-to-end framework that directly learns from raw point cloud observations and robot states, eliminating the need for pretrained models or external datasets. PRISM comprises three main components: a segmentation embedding unit that partitions the raw point cloud into distinct object clusters and encodes local geometric details; a cross-attention component that merges these visual features with processed robot joint states to highlight relevant targets; and a diffusion module that translates the fused representation into smooth robot actions. With training on 100 demonstrations per task, PRISM surpasses both 2D and 3D baseline policies in accuracy and efficiency within our simulated environments, demonstrating strong robustness in complex, object-dense scenarios. Code and some demos are available on https://github.com/czknuaa/PRISM.

Figures

Figures reproduced from arXiv: 2507.04633 by the authors.

Figure 1
Figure 1. PRISM is a visual imitation learning algorithm that marries 3D visual representations with diffusion policies, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Over view of PRISM architecture. PRISM takes a single-view 3D point cloud as input. First, it segments the point [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Cobot Magic platform in simulation environment [39]. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Attention heat map of Pick Apple Messy, Empty Cup Place Messy, Blocks Stack and Empty Cup Place tasks. Points [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Success Rate Trend across 6 Tasks. TABLE II: Ablation Experiment on Transformer Encoder Task PRISM (w/ TE) PRISM (w/o TE) Block Hammer Beat 96.4 90.0 Blocks Stack 80.2 75.2 Empty Cup Place 94.8 87.8 Empty Cup Place Messy 60.4 55.4 Mug Hanging 88.6 82.2 Pick Apple Messy…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 17 canonical work pages

  1. [1]

    Image-based visual servoing for enhanced cooperation of dual-arm manipulation,

    Z. Zhang, Y . Yang, W. Zuo, G. Song, A. Song, and Y . Shi, “Image-based visual servoing for enhanced cooperation of dual-arm manipulation,” IEEE Robotics and Automation Letters, 2025

    work page 2025

  2. [2]

    Spatialbot: Precise spatial understanding with vision language models,

    W. Cai, I. Ponomarenko, J. Yuan, X. Li, W. Yang, H. Dong, and B. Zhao, “Spatialbot: Precise spatial understanding with vision language models,” arXiv preprint arXiv:2406.13642, 2024

    arXiv 2024

  3. [3]

    Trends and challenges in robot manipulation,

    A. Billard and D. Kragic, “Trends and challenges in robot manipulation,” Science, vol. 364, no. 6446, p. eaat8414, 2019

    2019

  4. [4]

    High-speed autonomous robotic assembly using in-hand manipulation and re-grasping,

    T. Kang, J.-B. Yi, D. Song, and S.-J. Yi, “High-speed autonomous robotic assembly using in-hand manipulation and re-grasping,”Applied sciences, vol. 11, no. 1, p. 37, 2020

    work page 2020

  5. [5]

    Service robots in the healthcare sector,

    J. Holland, L. Kingston, C. McCarthy, E. Armstrong, P. O’Dwyer, F. Merz, and M. McConnell, “Service robots in the healthcare sector,” Robotics, vol. 10, no. 1, p. 47, 2021

    2021

  6. [6]

    A review of robot learning for manipulation: Challenges, representations, and algorithms,

    O. Kroemer, S. Niekum, and G. Konidaris, “A review of robot learning for manipulation: Challenges, representations, and algorithms,”Journal of machine learning research, vol. 22, no. 30, pp. 1–82, 2021

    2021

  7. [7]

    A survey on learning-based robotic grasping,

    K. Kleeberger, R. Bormann, W. Kraus, and M. F. Huber, “A survey on learning-based robotic grasping,”Current Robotics Reports, vol. 1, pp. 239–249, 2020

    work page 2020

  8. [8]

    Survey of imitation learning for robotic manipulation,

    B. Fang, S. Jia, D. Guo, M. Xu, S. Wen, and F. Sun, “Survey of imitation learning for robotic manipulation,”International Journal of Intelligent Robotics and Applications, vol. 3, no. 4, pp. 362–369, 2019

    work page 2019

Show all 41 references
  1. [9]

    A survey of imitation learning: Algorithms, recent developments, and challenges,

    M. Zare, P. M. Kebria, A. Khosravi, and S. Nahavandi, “A survey of imitation learning: Algorithms, recent developments, and challenges,” IEEE Transactions on Cybernetics, 2024

    2024

  2. [10]

    A reinforcement learning-based framework for robot manipulation skill acquisition,

    D. Liu, Z. Wang, B. Lu, M. Cong, H. Yu, and Q. Zou, “A reinforcement learning-based framework for robot manipulation skill acquisition,”IEEE Access, vol. 8, pp. 108429–108437, 2020

    2020

  3. [11]

    Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning,

    J. Hua, L. Zeng, G. Li, and Z. Ju, “Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning,”Sensors, vol. 21, no. 4, p. 1278, 2021

    2021

  4. [12]

    A survey on deep reinforcement learning algorithms for robotic manipulation,

    D. Han, B. Mulyana, V . Stankovic, and S. Cheng, “A survey on deep reinforcement learning algorithms for robotic manipulation,”Sensors, vol. 23, no. 7, p. 3762, 2023

    2023

  5. [13]

    Is behavior cloning all you need? understanding horizon in imitation learning,

    D. J. Foster, A. Block, and D. Misra, “Is behavior cloning all you need? understanding horizon in imitation learning,”arXiv preprint arXiv:2407.15007, 2024

    2024 arXiv

  6. [14]

    Behavioral cloning from observa- tion,

    F. Torabi, G. Warnell, and P. Stone, “Behavioral cloning from observa- tion,”arXiv preprint arXiv:1805.01954, 2018

    2018 arXiv

  7. [15]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on robot learning, pp. 158–168, PMLR, 2022

    2022

  8. [16]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    2023 arXiv

  9. [17]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023

    2023

  10. [18]

    Review of learning-based robotic manipulation in cluttered environments,

    M. Q. Mohammed, L. C. Kwek, S. C. Chua, A. Al-Dhaqm, S. Nahavandi, T. A. E. Eisa, M. F. Miskon, M. N. Al-Mhiqani, A. Ali, M. Abaker, et al., “Review of learning-based robotic manipulation in cluttered environments,”Sensors, vol. 22, no. 20, p. 7938, 2022

    2022

  11. [19]

    A systematic re- view of current and emergent manipulator control approaches,

    S. A. Ajwad, J. Iqbal, M. I. Ullah, and A. Mehmood, “A systematic re- view of current and emergent manipulator control approaches,”Frontiers of mechanical engineering, vol. 10, pp. 198–210, 2015

    2015

  12. [20]

    Generative adversarial imitation learning,

    J. Ho and S. Ermon, “Generative adversarial imitation learning,”Ad- vances in neural information processing systems, vol. 29, 2016

    2016

  13. [21]

    Act3d: 3d feature field transformers for multi-task robotic manipulation,

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki, “Act3d: 3d feature field transformers for multi-task robotic manipulation,” inConference on Robot Learning, pp. 3949–3965, PMLR, 2023

    2023

  14. [22]

    3d diffusion policy,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy,”arXiv e-prints, pp. arXiv–2403, 2024

    2024

  15. [23]

    Deep learning for 3d point clouds: A survey,

    Y . Guo, H. Wang, Q. Hu, H. Liu, L. Liu, and M. Bennamoun, “Deep learning for 3d point clouds: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 12, pp. 4338–4364, 2020

    2020

  16. [24]

    A review of deep learning-based semantic segmentation for point cloud,

    J. Zhang, X. Zhao, Z. Chen, and Z. Lu, “A review of deep learning-based semantic segmentation for point cloud,”IEEE access, vol. 7, pp. 179118– 179133, 2019

    2019

  17. [25]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 652– 660, 2017

    2017

  18. [26]

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

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Advances in neural information processing systems, vol. 30, 2017

    2017

  19. [27]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,”ACM Transactions on Graphics (tog), vol. 38, no. 5, pp. 1–12, 2019

    2019

  20. [28]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, X. Xu,et al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inkdd, vol. 96, pp. 226–231, 1996

    1996

  21. [29]

    Dbscan revisited, revisited: why and how you should (still) use dbscan,

    E. Schubert, J. Sander, M. Ester, H. P. Kriegel, and X. Xu, “Dbscan revisited, revisited: why and how you should (still) use dbscan,”ACM Transactions on Database Systems (TODS), vol. 42, no. 3, pp. 1–21, 2017

    2017

  22. [30]

    hdbscan: Hierarchical density based clustering.,

    L. McInnes, J. Healy, S. Astels,et al., “hdbscan: Hierarchical density based clustering.,”J. Open Source Softw., vol. 2, no. 11, p. 205, 2017

    2017

  23. [31]

    3d urban buildings extraction based on airborne lidar and photogrammetric point cloud fusion according to u-net deep learning model segmentation,

    P. Zhang, H. He, Y . Wang, Y . Liu, H. Lin, L. Guo, and W. Yang, “3d urban buildings extraction based on airborne lidar and photogrammetric point cloud fusion according to u-net deep learning model segmentation,” IEEE Access, vol. 10, pp. 20889–20897, 2022

    2022

  24. [32]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  25. [33]

    Neural machine translation by jointly learning to align and translate,

    D. Bahdanau, K. Cho, and Y . Bengio, “Neural machine translation by jointly learning to align and translate,”arXiv preprint arXiv:1409.0473, 2014

    2014 arXiv

  26. [34]

    Rethinking positional encoding in language pre-training,

    G. Ke, D. He, and T.-Y . Liu, “Rethinking positional encoding in language pre-training,”arXiv preprint arXiv:2006.15595, 2020

    2006 arXiv

  27. [35]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  28. [36]

    U-net and its variants for medical image segmentation: A review of theory and applications,

    N. Siddique, S. Paheding, C. P. Elkin, and V . Devabhaktuni, “U-net and its variants for medical image segmentation: A review of theory and applications,”IEEE access, vol. 9, pp. 82031–82057, 2021

    2021

  29. [37]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021

    2021

  30. [38]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, 2018

    2018

  31. [39]

    Robotwin: Dual-arm robot benchmark with generative digital twins (early version),

    Y . Mu, T. Chen, S. Peng, Z. Chen, Z. Gao, Y . Zou, L. Lin, Z. Xie, and P. Luo, “Robotwin: Dual-arm robot benchmark with generative digital twins (early version),”arXiv preprint arXiv:2409.02920, 2024

    2024 arXiv

  32. [40]

    Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations,

    T. Mu, Z. Ling, F. Xiang, D. Yang, X. Li, S. Tao, Z. Huang, Z. Jia, and H. Su, “Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations,”arXiv preprint arXiv:2107.14483, 2021

    2021 arXiv

  33. [41]

    Sapien: A simulated part-based interactive environment,

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang,et al., “Sapien: A simulated part-based interactive environment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11097–11107, 2020

    2020

Pith tools

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