acceptodds
Withdrawn from ICLR 2027

MultiPack: Hierarchical Planning with Physics Feedback for 3D Multi-Bin Packing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Packing household objects into storage bins requires deciding which objects to pack together, in what order, and in which poses. These decisions are tightly coupled: locally compact placements may increase overall bin usage, while physical settling can change the space available for subsequent objects. In this paper, we introduce , a hierarchical multi-bin packing framework that integrates cross-bin composition planning with physically validated in-bin pose search. Our Object–Group–Bin planning strategy groups objects by LLM-derived support, protection, and access priorities, connecting cross-bin allocation with in-bin scheduling. A capacity-based mixed-integer linear program assigns groups to bins by minimizing volume overflow. Within each bin, physics-feedback geometric search combines heightmap-guided candidate generation with height-envelope and weighted surface-clearance scoring. Physical validation, rollback, and settled-state synchronization close the loop between placements. Experiments on two datasets show reduced bin usage and competitive or improved layout quality under a shared evaluation protocol, while a real-world demonstration with 54 objects shows how the generated plans can guide human packing.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Irregular 3D Bin Packing, Multi-Bin Packing, Planning, Physics Feedback

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_skWfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies irregular 3D multi-bin packing of household objects when object geometry, names, and identical bins are given. MultiPack first asks an LLM to assign each object support, protection, and access priorities. Objects with similar attribute vectors are grouped, and a capacity-based mixed-integer linear program assigns groups to bins. Within each bin, objects are ordered according to the scheduling attributes and placed using heightmap-guided voxel search over preset orientations and discrete horizontal positions. Candidate actions are scored using height-envelope growth, surface clearance, and placement height. PyBullet simulation then checks descent, settling, penetration, and bin-boundary conditions. Accepted placements update the scene from the settled poses, while rejected trials are rolled back; unpacked objects can be reallocated later.

Critical strengths
  1. MultiPack connects bin composition with in-bin placement and physical state updates in one pipeline. This gives the framework a direct way to include scheduling preferences while coordinating decisions across bins.

  2. The rollback and settled-state synchronization are well aligned with rigid-body packing. Updating the voxel scene from all settled objects can account for motions caused by later placements, which is a meaningful design choice for dense scenes.

  3. The Object–Group–Bin hierarchy links cross-bin allocation with in-bin scheduling while reducing the MILP assignment variables from one per object–bin pair to one per group–bin pair. This is a clear way to keep the coarse planning stage manageable while preserving scheduling structure.

Critical weaknesses
  1. The LLM-derived attributes are only indirectly connected to physical feasibility. Support, protection, and access priorities do not encode shape compatibility, surface flatness, load-bearing ability, or actual retrieval paths. Objects with similar priority vectors can therefore remain geometrically incompatible. Groups are assigned as units within each planning round, so this coarse grouping can constrain allocation before unplaced objects are regrouped.

  2. The MILP objective is only a volume-overflow surrogate. Once each bin’s assigned volume is within the nominal capacity, all such allocations have zero objective, even when their group composition and geometric feasibility differ substantially. Geometric feasibility is deferred to later heuristics and simulation, so the cross-bin stage does not provide a principled optimization of the final packing objective.

  3. The in-bin search explores only preset orientations and discretized horizontal positions. A failed search can therefore indicate that the tested candidate set was insufficient rather than that no feasible continuous pose exists. Physical validation cannot recover feasible placements that candidate generation never proposes.

  4. The geometric score is a hand-designed proxy rather than a direct physical packing objective. Surface clearance is explicitly different from contact area or load-bearing capacity, and the height and clearance terms do not directly represent center of mass, protection from damage, or retrieval accessibility. The fixed weighting may consequently favor different placements across object categories and packing tasks.

  5. Physical acceptance is based on bounded simulator settling and geometric checks. The paper assumes known geometry and physical parameters and does not model robotic manipulation; its single 54-object demonstration shows that humans can follow the plans, but does not establish robust execution under uncertain object properties or grasp constraints.

  6. The multi-bin results include validated partial packings, exclude missing runs from averages, and count bins added during shared completion. However, the paper reports no per-method coverage or missing-run counts and does not separate native-planner bin counts from final counts, limiting interpretation of the comparison.

Questions for authors
  1. How often is the leftover-object reallocation stage triggered, and how often does it place objects that failed in their initial group assignments?

  2. When multiple group assignments yield the same volume-overflow objective, especially when the objective is zero, does the MILP use any secondary criterion or tie-breaking rule?

Reviewer_F3yHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed a 3D multi-bin packing approach, with the following major difference:

  • Items are grouped with LLM based on their semantic meaning. The items in the same group would be placed in the same box.
  • A looped mechanism that allows rollback when physical validation (i.e., considering the gravity of items) failed.
Critical strengths
  • This paper addressed an important optimization problem in real world (3D bin packing)
  • The idea of extracting the semantic information of items before packing is insightful. E.g., even with exactly the same size, a hardcover book and a pile of towers would be treated differently in real packing.
  • The figures in the paper are vivid (especially figure 3)
Critical weaknesses
  • The intuition behind LLM-derived scheduling attributes and grouping is not clear. More specifically,
    • Do we really need an LLM here? What about a simple pre-defined look-up table based on the the categories of items (which can be manually designed or assisted by AI)?
    • Why group similar items together? If we have to do some sort of grouping (maybe to reduce computational burden), a more intuitive way is to group items that are "complementary" with each other. Let's say, maximizing the packing efficiency within a group (total volume / the minimal volume of bounding box). In real world, there are also cases that we prefer not to group similar items. E.g., if I have two heavy machines, it might be more clever to divide them into separate boxes.
  • Apart from the usage of LLM in the grouping process (for which the need of LLM is questionable), this techniques used in the paper are generally standard heuristics or mathematical programming. The MILP model, heightmap, height-envelope and surface clearance scores, looped physical validation and rollback are not very new. Maybe their combination has some merits, but I feel like this paper would be more suitable for operational research venues to discover and recognize such merits.
Questions for authors

See "weaknesses"

Reviewer_LoGQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MultiPack, a framework for packing irregular 3D objects into multiple bins without packing-specific training. An LLM assigns support, protection, and accessibility priorities to guide object grouping and scheduling, while a MILP allocates groups to bins by minimizing volume overflow. Within each bin, geometric search generates candidate placements, and physics simulation validates them, with rollback for rejected trials and scene updates based on settled object poses. Unpacked objects are subsequently regrouped and reallocated. Experiments on two datasets report lower mean bin counts under a shared revalidation and completion protocol, while a 54-object human packing demonstration illustrates the framework’s practical use.

Critical strengths
  1. The Object–Group–Bin hierarchy provides a coherent framework connecting cross-bin allocation, object scheduling, and local placement.
  2. The method combines online physical validation, rollback, and settled-state updates to handle placement failures and object movement without requiring packing-specific policy training.
  3. Experiments across two datasets and multiple bin sizes show lower reported mean bin counts under the shared evaluation protocol, while ablations and a 54-object human packing demonstration provide additional evidence of the framework’s effectiveness and practical usability.
Critical weaknesses
  1. The MILP considers group volumes but not shape compatibility or physical packability. Low-overflow assignments may therefore remain unrealizable, while fixed grouping within each planning round can restrict otherwise effective object combinations.
  2. The evaluation is incomplete. PPN-Pack reports packed volume/count and planning time, while prior property-aware work such as OPA-Pack (Pan et al., 2025) explicitly measures protection outcomes; comparable metrics are missing here. Reporting object coverage would clarify whether fewer bins reflect equivalent task completion. Results before and after shared completion, together with rollback statistics, would also help establish the proposed components’ native contributions.
  3. MultiPack derives priorities solely from object names and task descriptions, whereas OPA-Pack also leverages visual and material cues. Without explicit load-bearing or retrieval models, these text-derived priorities may not reliably translate into physically protected or accessible layouts.
Questions for authors
  1. What drives the gains from physics feedback? Standard sequential baselines (e.g., PPN-Pack in simulation) already update the heightmap after each placement. Since the “No Physics Feedback” ablation removes online validation, rollback, and settled-state synchronization together, could the authors isolate the individual benefits of rollback/retry versus full-scene state synchronization of settling-induced shifts?
  2. Does group-level allocation itself improve packing? The object-level ablation changes both the allocation unit and the candidate-window scope (Sec 4.3). Does group-level allocation still reduce bin count when in-bin object ordering and candidate-selection rules are held constant?

If the authors can address most of my concerns, I would consider raising my score.

Reviewer_seTCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents an interesting problem of packing objects together in a box by taking the physical structure, scheduling preferences, and shape of the objects into account. The work focuses on a wide variety of different metrics and crafts an interesting objective function, which is further used to optimize packing efficiency.

The method includes multiple stages. (1) The first stage involves LLM-based grouping of different objects and subsequent MILP optimization for grouping the objects into bins. (2) Next, the objects in the bins undergo geometric optimization using heatmap-based search. (3) Finally, a physics simulator checks for any collisions and physically difficult configurations. An intersection of all these tests is required to be passed.

Critical strengths
  • The paper focuses on practical and important problems of geometric optimization. The problem is pretty relevant to real-world challenges.
  • The use of a physics simulator is novel and interesting. Furthermore, the work can be extended along this line to solve different optimization problems, including sensitivity to jerk, flips, etc.
  • The empirical evidence is extremely strong, and the paper presents compelling ablations of the methods. The results look fairly consistent and show a clear advantage over the existing baselines in both compactness metrics and reducing the number of boxes.
Critical weaknesses
  • The paper's novel contribution on the algorithmic side appears limited. The paper combines a vast majority of existing solutions to solve this problem in a rule-based manner. While the solution seems practical, it is unclear what the key algorithmic contribution beyond the combination of existing methods.. It would be good if the authors could go into more detail about the key algorithmic contributions apart from the integration of existing algorithms.
  • The efficiency of the optimization process is very unclear. It would be good if the authors could provide more details on the computational efficiency of this method compared to other baselines. It would be good to know how well it computational costs scale with the number of objects.
  • The paper relies on a lot of hyperparameters that are used without proper justification or empirical evidence. How sensitive are these methods to the hyperparameters, especially in the heatmap-guided search (Section 3.3)?
Questions for authors
  • One question is how the method handles and accounts for object-to-object compatibility. Two different objects with different materials can be clubbed together in a bin, or they might not be. For example, glass and rock, or a knife and a soft object. It would be interesting if MILP could be extended to account for object-to-object interactions as well.
  • Can the authors provide more insight into the simulation parameters and assumptions used? What kind of analyses are performed? It is understood that collision checks are performed, but more details would be helpful, such as whether compression of the objects was allowed or gravity was enabled, among a few other considerations.
  • Can the authors clarify why comparisons with the following reference papers [1] and [2] were not included, and how their approach differs from or improves upon these methods?

References

  1. Zhuang, Qiubing et al. “Dynamics simulation-based packing of irregular 3D objects.” Comput. Graph. 123 (2024): 103996.
  2. Zuo, Qiruyi et al. “The Three-dimensional Bin Packing Problem for Deformable Items.” 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM) (2022): 0911-0918.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.