Pith. sign in

REVIEW 3 major objections 3 minor 37 references

Online 3D Bin Packing with Fast Stability Validation and Stable Rearrangement Planning

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

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

Pith's one-line read The paper claims that checking only the newest item's center of gravity against a support polygon built from load-bearing top faces guarantees the entire packed bin stays stable, in near-constant time and without any item mass information.

desk verdict LBCP stability validation is broken by the height-less feasibility map, so the no-collapse guarantee is false; the rearrangement planner and runtime results still have some value. read the letter →

arxiv 2507.09123 v1 pith:VLGVFWVR submitted 2025-07-12 cs.RO

Ziyan Gao , Lijun Wang , Yuntao Kong , Nak Young Chong This is my paper · ORCID
classification cs.RO
keywords online3Dbinpackingload-bearableconvexpolygonstructuralstabilityvalidationrearrangementplanningMonteCarlotreesearchdeepreinforcementlearningrobotic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper targets a blind spot in online 3D bin packing: learned policies optimize space usage but cannot certify that a growing stack will not tip or collapse. It introduces the Load-Bearable Convex Polygon (LBCP), a horizontal region on an item's top face that can bear any vertical load at any point inside it, and argues that if a newly arriving item's center of gravity lies inside the support polygon formed by intersecting its footprint with existing LBCPs, then only that new item needs to be checked: the rest of the bin remains stable. Because the check uses only geometry plus a bounded uncertainty on the center of gravity, it needs no mass information and runs in nearly constant time, making it cheap enough to mask actions inside a deep reinforcement learning packing policy. When no stable direct placement exists, the companion Stable Rearrangement Planning module searches for a short sequence of unpack, pack, and repack moves that keeps the bin stable while freeing space. If the proof holds, the result is a packing controller that is both space-efficient and physically safe.

What carries the argument

The load-bearing object is the Load-Bearable Convex Polygon (LBCP), a horizontal convex region that can support any vertical load at any point inside it. Each packed item contributes one LBCP at its top face, and the bin floor contributes a special LBCP at height zero. The argument is carried by projecting all LBCPs onto a feasibility map aligned with a height map; the support polygon of a candidate placement is the convex hull of the intersection between the candidate's footprint and the feasible contact cells, and stability is decided by whether the candidate's bounded center-of-gravity uncertainty set lies inside that polygon. This map-based lookup makes validation cost nearly constant per candidate. A secondary mechanism is the Stable Rearrangement Planning module, which uses a Monte Carlo tree search restricted to unpacking moves to find which items must leave the bin, then an A* search over the precedence graph to shorten the operation sequence.

What would settle it

Build a three-layer stack: a large floor item, a middle item whose own center of gravity lies inside the floor item's footprint but whose top face extends beyond that support, and a new item placed on the overhanging part of the middle item's top face with its center of gravity over the overhang. The LBCP check would accept the new item because the middle item's full top face is treated as an LBCP; if the middle item tips or the stack collapses under the new load, the guarantee fails.

Watch

Extended reading notes

Core claim

The central discovery is that whole-stack stability can be certified locally. The authors define LBCPs as convex horizontal polygons that can support any gravitational force at any interior point, assigning one to each packed item's top face plus one to the bin floor. They prove that a cuboid resting stably on the floor has its entire top face as an LBCP, and that the convex hull of the intersections between a new item and existing LBCPs is itself an LBCP. It then follows, in their account, that whenever the new item's uncertain center of gravity lies fully inside that hull, the new item is stable, and the items supporting it remain stable as well. Consequently, the stability of the whole bin reduces to one geometric inclusion test on the newest item, executed by reading a precomputed feasibility map of LBCPs.

Load-bearing premise

The whole-stack guarantee rests on the assumption that every already packed item's entire top face is a Load-Bearable Convex Polygon; Lemma III.1 proves this only for items resting on the bin floor, while items in higher layers are assumed to inherit the property.

Editorial extensions

If this is right

  • A reinforcement learning packing policy can filter every candidate placement through the LBCP check, so the trained policy only ever proposes stable actions.
  • The stability check scales independently of bin size and item count, since it inspects one candidate against a lookup map rather than recomputing forces for the whole stack.
  • Because mass is never needed, the method transfers to heterogeneous boxes with unknown contents, as long as the center-of-gravity uncertainty bound is respected.
  • When no direct placement exists, the rearrangement plan remains stable at every intermediate step, and the sequence-refinement stage shortens the plan by roughly a third in the reported experiments.
  • The same stability module can be inserted into any placement proposal mechanism, not only the transformer-based deep reinforcement learning policy used in the paper.

Reading between the lines

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

  • The authors leave implicit that the guarantee is only as strong as the LBCP inventory: if higher-layer top faces are not fully load-bearing, the feasibility map over-approximates support, so a direct test would be to compare LBCP-approved placements against a rigid-body physics simulator on three-layer stacks.
  • Because the feasibility map is a simple geometric overlay, the same stability module could be bolted onto any placement generator, including classical heuristics, without retraining the policy.
  • The center-of-gravity uncertainty bound is a tunable safety margin; setting it too small maximizes utilization but weakens the no-collapse claim, so industrial deployments should set it from the measured worst-case box contents.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes an online 3D bin packing framework in which a learned DRL policy is constrained by a fast structural stability validation module and augmented by a Stable Rearrangement Planning (SRP) module with MCTS and A* search. The central theoretical contribution is the Load Bearable Convex Polygon (LBCP), a horizontal convex region that is claimed to support any gravitational force at any interior point. The paper claims that checking whether a new item's center of gravity falls inside a support polygon built from existing LBCPs guarantees that the whole bin never collapses, at nearly constant computational cost and without knowing item masses. The empirical sections report speedups over an adaptive stacking tree baseline, improved utilization with SRP, and a physical robot demonstration.

Significance. If the LBCP stability guarantee were correct, the proposed near-constant-time, mass-free stability check would be a meaningful advance for online bin packing and would justify the DRL integration and the SRP module. The paper also contains substantial engineering effort: a full DRL pipeline, a rearrangement planner, and a real-robot validation. However, the theoretical guarantee is the load-bearing contribution, and it is not established. The core proof covers only floor-supported items, and the algorithm's heightless feasibility map admits concrete counterexamples in which the method certifies an unstable stack. Since the central claim is unsound, the significance of the empirical results is reduced to heuristics without the advertised safety certificate.

major comments (3)
  1. [Section III-C2, Lemma III.1 and the sentence 'the number of LBCPs is equal to the number of packed items'] Lemma III.1 proves only that a cuboidal item resting directly on the bin floor has an entire top face that is an LBCP. It does not prove that an item placed on top of other items has a fully load-bearing top face. Yet the text states that every packed item contributes one LBCP, and Algorithm 2 marks the top face of every newly placed item as feasible. The proof of Corollary III.2.2 ('Due to the convexity of LBCP, the resultant... can be produced by Pt') is not a proof: it assumes that the support forces available at each LBCP can be chosen arbitrarily, which is exactly the property that must be established for non-floor items. Without upward propagation of load-bearing capacity, the stability of a multi-layer stack is not guaranteed by the proposed check.
  2. [Section III-C3, Algorithms 1 and 2] The feasibility map FM is a heightless projection of all LBCPs onto the xy-plane, and Algorithm 2 only adds points to FM and never removes covered lower LBCPs. In Algorithm 1, P_Sfeasible is any (x,y) with FM true, without verifying that the LBCP at that (x,y) is located at the computed support height hs. This allows lower-level LBCPs to validate a placement on an upper item's overhang. Concretely, let A be a 20x20x1 floor item, B a 10x10x1 item placed on A at (5,5)-(15,15), C a 12x12x1 item placed on B at (8,5)-(20,17), and D a 12x8x1 item placed on C at (12,8)-(20,16). D's center of gravity (16,12) lies inside the support polygon computed by Algorithm 1 because A's lower LBCP makes FM true under all of D's footprint, including C's unsupported overhang x>15. However, C is actually supported by B only up to x=15, and for m_D > m_C the combined center of gravity of C and D leaves C's true support polygon, so the stack tips. This directly contradicts Corollary III.2.1 and the abstract's claim of a 'guarantee no bin collapse', even under the paper's own static rigid-body assumptions.
  3. [Section III-C2, definition of LBCP and proof of Theorem III.2] The proof of Theorem III.2 is circular. An LBCP is defined as a polygon that can support any gravitational forces at any point within it. The theorem asserts that the convex hull of intersections with existing LBCPs is again an LBCP, with the proof saying that each intersected region can bear arbitrary forces and that the convex-hull operation guarantees convexity. Convexity alone does not imply that points in the hull that are not physically covered by any supporting region can bear loads; for example, the convex hull of two disjoint support patches includes the area between them, which has no material beneath it. Corollary III.2.1 then concludes stability directly from the definition of LBCP. The nontrivial content, namely that the support polygon is actually load-bearing, is assumed rather than proved. This circularity is not merely a presentation issue; it is the basis for the constant-time validation that omits mass information.
minor comments (3)
  1. [Section IV-A] The sentence 'any loading operation verified as stable by our method can ensure the bin structural stability' is an experimental overclaim, because the experiments measure only runtime, not stability, and the theoretical guarantee is not established.
  2. [Section III-D, Eq. (4)] The text contains a typo: 'The weight wvwe choose' should read 'The weight w_v we choose'.
  3. [Section III-C2, paragraph after Fig. 3] The discussion correctly identifies the failure of purely geometric support polygons in multi-layer stacks, but the proposed LBCP solution does not resolve that failure for non-floor items, as noted in the major comments.

Circularity Check

2 steps flagged · score 8.0 of 10

The no-collapse guarantee reduces to the definition of an LBCP: Corollary III.2.1's proof restates the defining property, and upper-layer LBCPs are assumed by the 'number of LBCPs equals number of packed items' assertion rather than derived.

  1. self definitional [Section III-C2, Theorem III.2 and Corollary III.2.1 (Eq. 2)]
    "According to Theorem III.2, the support polygon is considered an LBCP. From the LBCP definition, any point inside the support polygon can counteract any gravitational force. Thus, if the CoG of the newly added item Inew is located within the support polygon, the gravitational force will be balanced."

    LBCP is defined in the same section as 'characterized by its ability to support any gravitational forces at any point within it.' Corollary III.2.1's proof therefore does not derive stability from static equilibrium; it only applies the definition to the support polygon. The 'prediction' that a placement is stable is, by construction, the defining property of the polygon the algorithm constructs. Independent physical content is limited to Lemma III.1 (floor items); for upper layers the load-bearing property of the support polygon is inherited by definition, not established from forces, moments, or the supporting structure.

  2. self definitional [Section III-C2, paragraph before Lemma III.1; also Section III-C4/Algorithm 2 update]
    "The number of LBCPs increases with the growing number of packed items, and we use Pt to represent the set of LBCPs in the bin state Bt. Excluding (P△0 , 0), the number of LBCPs is equal to the number of packed items."

    This asserts that every packed item contributes an LBCP at its top face. Lemma III.1 proves only items on the floor. The inductive step would have to be Theorem III.2, but its proof ('From the definition of LBCPs... The operation CH guarantees convexity') again rests on the definition. So the assumption that each upper item's top face is fully load-bearing is exactly the conclusion needed for the no-collapse guarantee; it is imported into the input set Pt rather than derived. Algorithm 2 compounds this by adding the new support polygon to the 2D feasibility map without ever removing lower covered LBCPs, so height information is lost and lower-level LBCPs are treated as validating upper placements in Alg. 1 lines 11-14.

full rationale

The central stability claim is not derived from independent rigid-body statics. The paper defines an LBCP as a polygon that 'can support any gravitational forces at any point within it,' then proves Corollary III.2.1 by saying the support polygon is an LBCP and therefore any CoG inside it is balanced. That is a definitional restatement. Lemma III.1 provides a genuine base case for floor items, but the extension to stacks relies on Theorem III.2, whose proof is one sentence invoking the LBCP definition and convexity; it does not show that the convex hull of load-bearing regions is load-bearing in the sense required, nor that a stable item's whole top face becomes load-bearing. The assertion that the number of LBCPs equals the number of packed items builds the unproven upper-layer property into the input state. A concrete failure mode is visible in the algorithm: FM is a height-less projection of all LBCPs and Algorithm 2 only sets entries to true, so a lower item's LBCP can mark a point feasible at an upper support height; the counterexample given in the skeptic analysis (A floor, B on A, C overhanging B, heavy D on C's overhang) violates Corollary III.2.1 even under the paper's idealized rigid-body assumptions. This is more than a minor self-citation issue: it is the main guarantee of the paper. However, the paper does contain independent elements (floor-item lemma, standard support-polygon geometry, benchmarks), so a score of 10 would overstate; a score of 8 reflects that the headline no-collapse result is forced by definition for all non-floor layers and is not validated by an independent equilibrium derivation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central claim relies on the LBCP construct, whose load-bearing property is assumed for all packed items except floor-level ones. Additional parameters and search limits control the rearrangement planner's behavior and reported utilization gains. The paper does not justify the upward propagation of the LBCP property, which is the main unsupported premise.

free parameters (6)
  • CoG uncertainty bound deltaCoG = not specified in paper
    Defines the uncertainty box for each item's center of gravity (Eq. 1). The stability check validates all extreme points of this box, so the bound directly controls which placements pass.
  • MCTS exploration weight eta = 1
    Set in Eq. 3 for UCB1; affects which unpacking candidates are explored in SRP.
  • Rollout reward weight w_v = 5
    Weights critic value against immediate utilization in Eq. 4 for MCTS rollouts.
  • Target utilization T_uti = 0.8
    Stops rearrangement search once bin utilization reaches this threshold; directly affects reported utilization gain.
  • MCTS search limits = max nodes 100, depth 6, branching 3
    These caps control success rate and plan length; the reported 83.9% success rate is specific to these values.
  • Staging capacity = 6 in simulation, 4 in real experiment
    Limits the number of unpacked items; influences operation count and feasibility.
assumptions (5)
  • domain assumption Items are cuboids and placements are axis-aligned
    Sets the geometry for support regions and LBCPs throughout Section III.
  • standard math An item is stable iff its CoG lies within its support polygon
    Adopted from Ramos et al. [27] in Section III-C1; ignores friction, torque balance, and dynamic effects.
  • ad hoc to paper Every packed item's top face is an LBCP
    Stated in Section III-C2 ('the number of LBCPs is equal to the number of packed items'); Lemma III.1 only proves the floor-item case, so this upward propagation is assumed.
  • domain assumption CoG uncertainty is bounded by deltaCoG proportionally to item dimensions
    Eq. 1; the paper assumes this bounded uncertainty model without empirical justification.
  • ad hoc to paper Support forces at any point inside an LBCP can have arbitrary magnitude
    This is the defining property of an LBCP; it is assumed rather than derived, and is used in Corollary III.2.2 to argue that supporting items remain stable.
invented entities (2)
  • Load Bearable Convex Polygon (LBCP) independent evidence
    purpose: A polygon on the top face of each packed item that is asserted to bear any vertical load at any interior point, used to validate stability without masses.
    The paper's real-robot video is a falsifiable check, but the experiments do not measure failure rates or compare against physics simulation; the property itself is assumed for upper items.
  • Feasibility map (FM)
    purpose: A grid projection of all LBCPs onto the xy-plane so support-polygon computation reduces to Boolean lookups.
    A data structure internal to the algorithm; it has no independent physical meaning and its correctness depends on the LBCP assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online 3D Bin Packing with Fast Stability Validation and Stable Rearrangement Planning." pith.science (2026). https://pith.science/paper/VLGVFWVR

@misc{pith2026250709123,
  author       = {Pith},
  title        = {Pith review of: Online 3D Bin Packing with Fast Stability Validation and Stable Rearrangement Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLGVFWVR}},
  note         = {Machine review of arXiv:2507.09123}
}
read the original abstract

The Online Bin Packing Problem (OBPP) is a sequential decision-making task in which each item must be placed immediately upon arrival, with no knowledge of future arrivals. Although recent deep-reinforcement-learning methods achieve superior volume utilization compared with classical heuristics, the learned policies cannot ensure the structural stability of the bin and lack mechanisms for safely reconfiguring the bin when a new item cannot be placed directly. In this work, we propose a novel framework that integrates packing policy with structural stability validation and heuristic planning to overcome these limitations. Specifically, we introduce the concept of Load Bearable Convex Polygon (LBCP), which provides a computationally efficient way to identify stable loading positions that guarantee no bin collapse. Additionally, we present Stable Rearrangement Planning (SRP), a module that rearranges existing items to accommodate new ones while maintaining overall stability. Extensive experiments on standard OBPP benchmarks demonstrate the efficiency and generalizability of our LBCP-based stability validation, as well as the superiority of SRP in finding the effort-saving rearrangement plans. Our method offers a robust and practical solution for automated packing in real-world industrial and logistics applications.

Figures

Figures reproduced from arXiv: 2507.09123 by the authors.

Figure 1
Figure 1. The proposed online bin packing framework consists of three key modules: the structural stability module (Sections III-C2 and III-C3), the packing policy [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Support regions and support polygon. Item [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Load unbearable region in the support polygon. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Precedence graphs before and after re-arrangement. Note that [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Use of A∗ for finding the shortest sequence length of operations. Nodes represent the bin state and the set of unpacked items, while edges represent the operations. The numbers in red are heuristic costs. b cannot be picked up due to the movement block (MB) of item a. …
Figure 7
Figure 7. Figure 7: Time cost for completing the validation of all pre-specified loading positions for the new item. Each bar is annotated with the mean and standard [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Performance of the trained DRL model on the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Joint distribution of the number of operations and bin utilization. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Success rate of the MCTS for finding the feasible sequence [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Real packing test with cardboard boxes of various sizes. The first two snapshots illustrate the sequential loading of new items, while the subsequent [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

  1. [1]

    Multidimensional bin packing and other related problems: A survey,

    H. I. Christensen, A. Khan, S. Pokutta, and P. Tetali, “Multidimensional bin packing and other related problems: A survey,” Computer Science Review, 2016

    work page 2016

  2. [2]

    Online 3d bin packing with constrained deep reinforcement learning,

    H. Zhao, Q. She, C. Zhu, Y . Yang, and K. Xu, “Online 3d bin packing with constrained deep reinforcement learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 1, 2021, pp. 741–749

    work page 2021

  3. [3]

    Learning efficient online 3d bin packing on packing configuration trees,

    H. Zhao, Y . Yu, and K. Xu, “Learning efficient online 3d bin packing on packing configuration trees,” in International conference on learning representations, 2021

    work page 2021

  4. [4]

    Packerbot: Variable-sized product packing with heuristic deep reinforcement learning,

    Z. Yang, S. Yang, S. Song, W. Zhang, R. Song, J. Cheng, and Y . Li, “Packerbot: Variable-sized product packing with heuristic deep reinforcement learning,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 5002–5008

    work page 2021

  5. [5]

    Learning practically feasible policies for online 3d bin packing,

    H. Zhao, C. Zhu, X. Xu, H. Huang, and K. Xu, “Learning practically feasible policies for online 3d bin packing,” Science China Information Sciences, vol. 65, no. 1, p. 112105, 2022

    work page 2022

  6. [6]

    Gopt: Generalizable online 3d bin packing via transformer-based deep reinforcement learning,

    H. Xiong, C. Guo, J. Peng, K. Ding, W. Chen, X. Qiu, L. Bai, and J. Xu, “Gopt: Generalizable online 3d bin packing via transformer-based deep reinforcement learning,” IEEE Robotics and Automation Letters , vol. 9, no. 11, pp. 10 335–10 342, 2024

    work page 2024

  7. [7]

    Efficient reinforcement learning of task planners for robotic palletization through iterative action masking learning,

    Z. Wu, Y . Li, W. Zhan, C. Liu, Y .-H. Liu, and M. Tomizuka, “Efficient reinforcement learning of task planners for robotic palletization through iterative action masking learning,” IEEE Robotics and Automation Letters, 2024. IEEE Journal Paper, VOL. XX, NO. XX, XXXX 10 Fig. 11. Real packing test with cardboard boxes of various sizes. The first two snapsh...

    work page 2024

  8. [8]

    Gradual receptive expansion using vision transformer for online 3d bin packing,

    M. Kang, H. Kee, Y . Park, J. Kim, J. Jeong, G. Cheon, J. Lee, and S. Oh, “Gradual receptive expansion using vision transformer for online 3d bin packing,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 7338–7343

    work page 2024

Show all 37 references
  1. [9]

    3d vision robot online packing platform for deep reinforcement learning,

    X. Mu, Q. Kan, Y . Jiang, C. Chang, X. Tian, L. Zhou, and Y . Zhao, “3d vision robot online packing platform for deep reinforcement learning,” Robotics and Computer-Integrated Manufacturing , vol. 94, p. 102941, 2025

    2025

  2. [10]

    Tap-net: transport-and-pack using reinforcement learning,

    R. Hu, J. Xu, B. Chen, M. Gong, H. Zhang, and H. Huang, “Tap-net: transport-and-pack using reinforcement learning,” ACM Transactions on Graphics (TOG), vol. 39, no. 6, pp. 1–15, 2020

    2020

  3. [11]

    Neural packing: from visual sensing to reinforcement learning,

    J. Xu, M. Gong, H. Zhang, H. Huang, and R. Hu, “Neural packing: from visual sensing to reinforcement learning,” ACM Transactions on Graphics (TOG), vol. 42, no. 6, pp. 1–11, 2023

    2023

  4. [12]

    Towards online 3d bin packing: Learning synergies between packing and unpacking via drl,

    S. Song, S. Yang, R. Song, S. Chu, W. Zhang et al. , “Towards online 3d bin packing: Learning synergies between packing and unpacking via drl,” in Conference on Robot Learning . PMLR, 2023, pp. 1136–1145

    2023

  5. [13]

    Heuristics integrated deep reinforcement learning for online 3d bin packing,

    S. Yang, S. Song, S. Chu, R. Song, J. Cheng, Y . Li, and W. Zhang, “Heuristics integrated deep reinforcement learning for online 3d bin packing,” IEEE Transactions on Automation Science and Engineering , vol. 21, no. 1, pp. 939–950, 2023

    2023

  6. [14]

    Attend2pack: Bin packing through deep reinforcement learning with attention,

    J. Zhang, B. Zi, and X. Ge, “Attend2pack: Bin packing through deep reinforcement learning with attention,”arXiv preprint arXiv:2107.04333, 2021

    2021 arXiv

  7. [15]

    Learning to pack: A data-driven tree search algorithm for large-scale 3d bin packing problem,

    Q. Zhu, X. Li, Z. Zhang, Z. Luo, X. Tong, M. Yuan, and J. Zeng, “Learning to pack: A data-driven tree search algorithm for large-scale 3d bin packing problem,” in Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 2021, pp. 4393– 4402

    2021

  8. [16]

    Learning to solve 3-d bin packing problem via deep reinforcement learning and constraint programming,

    Y . Jiang, Z. Cao, and J. Zhang, “Learning to solve 3-d bin packing problem via deep reinforcement learning and constraint programming,” IEEE transactions on cybernetics , vol. 53, no. 5, pp. 2864–2875, 2021

    2021

  9. [17]

    Solving 3d packing problem using transformer network and reinforcement learning,

    Q. Que, F. Yang, and D. Zhang, “Solving 3d packing problem using transformer network and reinforcement learning,” Expert Systems with Applications, vol. 214, p. 119153, 2023

    2023

  10. [18]

    Ppn-pack: Placement proposal network for efficient robotic bin packing,

    J.-H. Pan, X. Gao, K.-H. Hui, S. Zhu, Y .-H. Liu, P.-A. Heng, and C.- W. Fu, “Ppn-pack: Placement proposal network for efficient robotic bin packing,” IEEE Robotics and Automation Letters , 2024

    2024

  11. [19]

    Bin packing optimization via deep reinforcement learning,

    B. Wang, Z. Lin, W. Kong, and H. Dong, “Bin packing optimization via deep reinforcement learning,” IEEE Robotics and Automation Letters , vol. 10, no. 3, pp. 2542–2549, 2025

    2025

  12. [20]

    Stable bin packing of non-convex 3d objects with a robot manipulator,

    F. Wang and K. Hauser, “Stable bin packing of non-convex 3d objects with a robot manipulator,” in 2019 International Conference on Robotics and Automation (ICRA) , 2019, pp. 8698–8704

    2019

  13. [21]

    Stablelego: Stability analysis of block stacking assembly,

    R. Liu, K. Deng, Z. Wang, and C. Liu, “Stablelego: Stability analysis of block stacking assembly,” IEEE Robotics and Automation Letters , vol. 9, no. 11, pp. 9383–9390, 2024

    2024

  14. [22]

    Efficient selectivity and backup operators in monte-carlo tree search,

    R. Coulom, “Efficient selectivity and backup operators in monte-carlo tree search,” in International conference on computers and games . Springer, 2006, pp. 72–83

    2006

  15. [23]

    A formal basis for the heuristic determination of minimum cost paths,

    P. Hart, N. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” IEEE Transactions on Systems Science and Cybernetics , vol. 4, no. 2, pp. 100–107, 1968. [Online]. Available: https://doi.org/10.1109/tssc.1968.300136

    1968

  16. [24]

    An introduction to convolutional neural networks,

    K. O’Shea and R. Nash, “An introduction to convolutional neural networks,” 2015. [Online]. Available: https://arxiv.org/abs/1511.08458

    2015 arXiv

  17. [25]

    A maximal-space algorithm for the container loading problem,

    F. Parre ˜no, R. Alvarez-Vald ´es, J. M. Tamarit, and J. F. Oliveira, “A maximal-space algorithm for the container loading problem,” INFORMS Journal on Computing , vol. 20, no. 3, pp. 412–422, 2008

    2008

  18. [26]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garn...

    2017

  19. [27]

    A physical packing se- quence algorithm for the container loading problem with static mechan- ical equilibrium conditions,

    A. G. Ramos, J. F. Oliveira, and M. P. Lopes, “A physical packing se- quence algorithm for the container loading problem with static mechan- ical equilibrium conditions,” International Transactions in Operational Research, vol. 23, no. 1-2, pp. 215–238, 2016

    2016

  20. [28]

    The pallet loading problem: Three-dimensional bin packing with practical constraints,

    F. Gzara, S. Elhedhli, and B. C. Yildiz, “The pallet loading problem: Three-dimensional bin packing with practical constraints,” European Journal of Operational Research, vol. 287, no. 3, pp. 1062–1074, 2020

    2020

  21. [29]

    Solving a 3d bin packing problem with stacking constraints,

    W. Zhu, S. Chen, M. Dai, and J. Tao, “Solving a 3d bin packing problem with stacking constraints,” Computers & Industrial Engineering, vol. 188, p. 109814, 2024. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0360835223008380

    2024

  22. [30]

    Static stability versus packing efficiency in online three-dimensional packing problems: A new ap- proach and a computational study,

    S. Ali, A. G. Ramos, and J. F. Oliveira, “Static stability versus packing efficiency in online three-dimensional packing problems: A new ap- proach and a computational study,” Computers & Operations Research , p. 107005, 2025

    2025

  23. [31]

    An efficient deep reinforce- ment learning model for online 3d bin packing combining object rear- rangement and stable placement,

    P. Zhou, Z. Gao, C. Li, and N. Y . Chong, “An efficient deep reinforce- ment learning model for online 3d bin packing combining object rear- rangement and stable placement,” in 2024 24th International Conference on Control, Automation and Systems (ICCAS) , 2024, pp. 964–969

    2024

  24. [32]

    Efficient reinforcement learning of task planners for robotic palletization through iterative action masking learning,

    Z. Wu, Y . Li, W. Zhan, C. Liu, Y .-H. Liu, and M. Tomizuka, “Efficient reinforcement learning of task planners for robotic palletization through iterative action masking learning,” IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9303–9310, 2024

    2024

  25. [33]

    Physics-aware robotic palletization with online masking inference,

    T. Zhang, Z. Wu, Y . Chen, Y . Wang, B. Liang, S. Moura, M. Tomizuka, M. Ding, and W. Zhan, “Physics-aware robotic palletization with online masking inference,” arXiv preprint arXiv:2502.13443 , 2025

    2025 arXiv

  26. [34]

    Physical question, virtual answer: Optimized real-time physical simulations and physics- informed learning approaches for cargo loading stability,

    P. G. Mazur, J. W. Melsbach, and D. Schoder, “Physical question, virtual answer: Optimized real-time physical simulations and physics- informed learning approaches for cargo loading stability,” Operations Research Perspectives, vol. 14, p. 100329, 2025. [Online]. Available: ht...

    2025

  27. [35]

    Finite-time analysis of the multiarmed bandit problem,

    P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the multiarmed bandit problem,” Machine Learning , vol. 47, no. 2-3, pp. 235–256, 2002. [Online]. Available: https://link.springer.com/article/10. 1023/A:1013689704352

    2002

  28. [36]

    Tianshou: A highly modularized deep reinforcement learning library,

    J. Weng, H. Chen, D. Yan, K. You, A. Duburcq, M. Zhang, Y . Su, H. Su, and J. Zhu, “Tianshou: A highly modularized deep reinforcement learning library,” Journal of Machine Learning Research , vol. 23, no. 267, pp. 1–6, 2022

    2022

  29. [37]

    Automatic generation and detection of highly reliable fiducial markers under occlusion,

    S. Garrido-Jurado, R. Mu ˜noz-Salinas, F. J. Madrid-Cuevas, and M. J. Mar´ın-Jim´enez, “Automatic generation and detection of highly reliable fiducial markers under occlusion,” Pattern Recognition, vol. 47, no. 6, pp. 2280–2292, 2014

    2014

Pith tools

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