Pith. sign in

REVIEW 3 major objections 5 minor 6 references

PackingGPT: 3D Packing Agent for Real Furniture in Last-Mile Delivery

T0 review · 3 major / 5 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 fine-tuning a 1-billion-parameter language model on centre-of-mass-gated packing sequences reduces per-box instability in furniture-in-vehicle packing to 0.67% on an SUV benchmark and 3.3% on a sedan benchmark, while g

desk verdict New dataset and an LLM packing idea that deserve a look, but the central comparison fails because the baselines never see the stability constraint. read the letter →

arxiv 2608.01427 v1 pith:HE5AVTCQ submitted 2026-08-02 cs.CV

Yi You , Hui Li This is my paper
classification cs.CV
keywords 3Dbinpackingcentre-of-massstabilityflat-packfurniturelast-miledeliverylargelanguagemodelautoregressiveplacementvoxeloccupancyheterogeneous
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 — titled PackingGPT and developed in the text as the Lego-Language Packing (LLP) framework — is trying to establish that the physically stable packing of heterogeneous flat-pack furniture boxes into realistic, partially occupied vehicle trunks can be learned by a small language model rather than hand-coded container-loading heuristics. It builds a real flat-pack packaging dataset, generates training sequences in which every placement passes a centre-of-mass (CoM) stability gate, and fine-tunes LLaMA-3.2-1B to emit box ID, coordinates, and rotation one token at a time. On 500 SUV scenarios the model yields 0.67% per-box CoM failures, and on 500 sedan scenarios 3.3%, against 11.3–34.2% for five sorting heuristics that ignore CoM. The authors are explicit that the current setting considers only balance, allows six axis-aligned orientations, and ignores boxes that cannot be stacked on top; the claim is therefore about axis-aligned, balance-only packing. If the claim holds, pre-trip feasibility checks for last-mile delivery and customer self-pickup become practical, and the same token-placement format extends to new products and clutter layouts without retraining.

What carries the argument

The central object is the CoM-gated packing sequence, generated by a 'Lego disassembly' data pipeline and consumed by a causal language model. Formally, a training example is the concatenation $\mathcal{T}_t = \mathcal{P} \circ \mathrm{shuffle}(\mathcal{I}) \circ \mathcal{S} \circ \mathcal{O}$, where $\mathcal{P}$ is the vehicle/catalogue prompt, $\mathcal{I}$ the unordered box IDs, $\mathcal{S}$ a separator, and $\mathcal{O}$ the ordered output placements $(\mathrm{ID}_i, x_i, y_i, z_i, \theta_i)$. Only output tokens contribute to the loss. The load-bearing stability predicate is the four-condition CoM gate: support existence; rejection of degenerate 1–2 point contacts for boxes above 5 cm;

What would settle it

Run the exact SUV-500 and Sedan-500 packings from the trained model through a rigid-body physics solver (or an instrumented trunk) under measured city-driving accelerations and count boxes that tip or slide. If the instability rate is far above 0.67% / 3.3%, the CoM gate is not a faithful stability model. Alternatively, replace the voxel-cell-centre support set with true contact polygons and recompute the four stability conditions: if many accepted placements become rejected, the metric depends on voxel discretisation.

Watch

Extended reading notes

Core claim

PackingGPT — the Lego-Language Packing (LLP) framework — treats furniture loading as autoregressive token generation. A fine-tuned LLaMA-3.2-1B-Instruct reads a prompt listing the vehicle dimensions, the box catalogue, and the unordered order, then outputs for each box its ID, corner coordinates, and one of six axis-aligned rotations. Training sequences are produced by a reverse-Lego data generator: a cluttered voxelised trunk is unpacked top-down, then boxes are re-placed bottom-up into the remaining gaps, with every committed placement required to pass a centre-of-mass gate. The gate rejects support-free placements, degenerate point/line contacts for boxes taller than 5 cm, placements whos

Load-bearing premise

The load-bearing premise is that the four-condition centre-of-mass checker — a box is stable iff its geometric-centre projection lies inside the convex hull of its supporting voxels with a $0.15$-height anti-tipping margin, and floor boxes are unconditionally stable — matches real tipping in a moving vehicle; the same checker produces every training label and every reported failure rate, so if it does not match reality the headline numbers are simulator artifacts.

Editorial extensions

If this is right

  • Pre-trip feasibility checks become practical: given a customer order and a vehicle, the model predicts both fit and per-box stability, supporting last-mile route planning and customer self-pickup.
  • The policy generalises to unseen products and changed cargo layouts without retraining, so one model can serve varied orders and vehicle configurations.
  • Deployment to a new environment is cheap: on Sedan-500, training on 100 scenarios already drops the CoM failure rate below 2%.
  • Because the language model alone places slightly fewer boxes than the best heuristics, the immediate practical recipe is a hybrid: LLM-generated stable order plus constrained decoding, collision-checked rejection sampling, or a heuristic.
  • New operational constraints, such as crush-sensitive boxes, unloading order, and axle-weight limits, can be added by changing the structured tokens in the prompt rather than retraining the model.

Reading between the lines

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

  • Because the same CoM checker produces both the training labels and the evaluation metric, the 0.67% and 3.3% numbers measure agreement with that checker, not with real vehicle physics; certifying real-road stability would require a physics-based or learned stability oracle.
  • The tokenised placement format should transfer to other constrained spatial assembly tasks, such as palletising or mixed-freight van loading, where new constraints can be written as additional structured tokens.
  • The utilisation figures (roughly 54% SUV, 46% sedan) at near-zero stability failure suggest that in last-mile furniture delivery the binding constraint is placement order and stability, not raw space; a practical system could therefore recommend an order rather than only a final arrangement.
  • A direct testable extension is to compare LLP with a reinforcement-learning packing policy trained on the same CoM-gated benchmark; the paper motivates this comparison but does not run it.
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 / 5 minor

Summary. The paper proposes PackingGPT/LLP: a fine-tuned LLaMA-3.2-1B model that treats 3D furniture packing as autoregressive token generation. It introduces a dataset of 202 real IKEA flat-pack products (300 boxes), five BLF baselines, and two simulated vehicle-trunk benchmarks (SUV-500, Sedan-500) with center-of-mass (CoM) stability constraints. The headline claim is that LLP reduces per-box CoM failure to 0.67% (SUV) and 3.3% (Sedan), versus 10–40% for baselines, and generalizes to unseen products and clutter layouts without fine-tuning. The paper also reports a training-data ablation and visualizations of selected packing sequences.

Significance. If the central comparison were clean, this would be a useful demonstration that a small language model can act as a packing policy under stability constraints, and the real flat-pack dataset is a genuine contribution to a data-scarce logistics problem. The tokenization scheme and fine-tuning cost (about 14 minutes on one GPU) are practical. However, the main empirical claim is not currently established. The baselines are run without CoM constraints while LLP is trained and evaluated with the same CoM-gated checker, making the failure-rate gap at least partly an artifact of differential access to the objective. The manuscript also contains an internal arithmetic inconsistency in a key table. With proper control experiments and corrected reporting, the contributions could form a solid benchmark paper.

major comments (3)
  1. [§2.3 Baseline methods; Tables 3–4; Abstract] The headline comparison is confounded. The five baselines are explicitly run 'without considering the CoM constraints,' while LLP is trained on CoM-gated sequences and evaluated against the same stability checker. The paper itself shows the decisive effect of the gate: on the example set, with CoM used as a hard gate, Large-First, Longest-First, and Random-Best 'achieve full occupancy with zero stability failures' (§2.3). No CoM-gated baseline numbers are reported on SUV-500/Sedan-500. Without those runs, the 0.67% versus 18.9–34.2% gap measures access to the checker rather than model quality. Please add the five baselines with the same CoM gate on the 500-scenario eval sets, and discuss whether CoM-compliant orderings exist there as they do in the example set.
  2. [LEGO Format Data Generation; Dataset Settings; Table 2] Training labels and the evaluation metric are produced by the same stability routine. The stability definition is ad hoc: a box is stable iff the geometric-center projection is inside the convex hull of support points with margin alpha = 0.15, with floor boxes unconditionally stable; the generator also uses a 35% support-ratio pre-filter and 1-cm voxel checks. All failure rates are therefore internal to this simulator. To support the physical-stability claim, the paper should report sensitivity of the reported rates to alpha and to voxel resolution, and compare a subset of placements against a physics-based surrogate or human labels. If the checker does not reflect real tipping in a moving vehicle, the 0.67% figure does not transfer to the claimed real-world setting.
  3. [Table 5] The failure-rate column in Table 5 is arithmetically wrong by an order of magnitude. For the rows: 42/2274 = 1.85% (not 0.18%), 66/2283 = 2.89% (not 0.28%), 50/2318 = 2.16% (not 0.21%), 78/2350 = 3.32% (not 0.33%), and 78/2397 = 3.25% (not 0.32%). The corrected 500-scenario value (3.25%) is consistent with the 3.3% in Table 4, and the corrected 100-scenario value (1.85%) still supports the 'below 2%' claim, but the table as printed is unusable. Please correct the table and reconcile the accompanying text.
minor comments (5)
  1. [Section 2.3 numbering] Three subsections are numbered 2.3: 'Center-of-Mass,' 'Baseline methods,' and 'Lego-Language Packing.' Renumber the sections.
  2. [Abstract and text] The abstract refers to 'LTP' while the body uses 'LLP'; the Lego terminology is also inconsistent ('Lego' vs 'LEGO'). Please unify.
  3. [Table 1] Several cells are blank (dimension means for Tier 3) and the volume means of 0.911 and 0.385 are implausible in liters; these are likely cubic meters. Clarify units and fill missing entries.
  4. [Throughout] There are informal/noisy phrases such as 'Ithe box is on the floor,' 'and e.t.,' and 'overcapacity' in Figure 4(b). A careful editing pass is needed.
  5. [Reproducibility] Please provide a link to the dataset and code. The TidyTuesday reference URL should be verified ('tidytues.day' appears to be missing a 't').

Circularity Check

1 steps flagged · score 6.0 of 10

Reported CoM failure-rate reduction is driven by the CoM gate shared by training and evaluation, while baselines were denied the gate.

  1. fitted input called prediction [Abstract; Section 2.3 (Center-of-Mass and Baseline methods); Tables 3 and 4]
    "Five baseline packing methods were tested on our dataset without considering the Centre-of-Mass (CoM) constraints. In sedan car simulations, 10-40% of placed boxes failed the stability check on average. When the LLP model was trained on packing sequences with CoM constraints enforced during placement, the failure rate dropped to 0.67% (SUV-500)."

    The headline improvement is an artifact of the experimental split: baselines are explicitly run 'without considering the CoM constraints,' while LLP is trained on sequences that are 'CoM-gated' (Table 2: 'CoM compliance is 100% by construction') and evaluated with the same four-condition CoM checker used to create the training labels. The paper itself shows (§2.3) that when the CoM gate is applied to Large-First, Longest-First, and Random-Best, they 'achieve full occupancy with zero stability failures.' Thus the 10-40% vs 0.67% gap is largely the effect of switching on the checker, not evidence that the learned policy is physically superior. The evaluation metric is the same filter that generated the training data, so the low failure rate mostly reflects imitation of that filter, and no ex

full rationale

The central circular step is the shared CoM checker: it generates the training labels, filters the training placements, and defines the evaluation metric, while the baseline methods are deliberately run without it. The paper's own demonstration that CoM-gated heuristics achieve zero failures on the example set confirms that the reported reduction is attributable to the gate itself rather than to the learned model. The held-out evaluation on unseen products and cargo configurations provides some genuine generalization evidence, so the circularity is partial rather than total; however, that held-out set is still scored by the same checker that produced the training data. No self-citation load-bearing arguments or imported uniqueness theorems appear. The appropriate corrective would be to report CoM-gated baseline results on SUV-500/Sedan-500 and to validate the checker against real tipping behavior, which would break the self-referential loop.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central numbers depend on several hand-set thresholds (alpha, voxel resolutions, support ratio, inventory limits) and on the validity of a purely geometric stability checker. No new physical entities are introduced. The dataset itself is collected real-world data, not an invented entity.

free parameters (3)
  • anti-tipping margin coefficient alpha = 0.15
    Chosen as 'a conservative ratio' in Section 2.3(d), not calibrated against physical experiments. It directly determines which placements are accepted in training data and evaluation.
  • voxel resolutions and support-ratio pre-filter threshold = 5 cm pre-filter, 1 cm precise check, 35% support ratio
    Hand-chosen in Section 'LEGO Format Data Generation'. These thresholds shape the set of feasible placements and therefore the learned policy.
  • randomized inventory limits = large objects 1 to 3, small objects up to 10
    Section 'Randomised box inventory' uses these ranges to simulate human behavior; the distribution of training scenarios depends on them.
assumptions (5)
  • domain assumption Vehicle trunks can be modeled as rectangular cuboids with clutter as fixed voxels; wheel wells, seat folds, and irregular contours are ignored.
    Section 1.3 explicitly abstracts away these geometries, so the simulator operates on a simplified cargo space.
  • domain assumption Each flat-pack box is rigid, axis-aligned, and can take only six orientations; tilting is not allowed.
    Section 2.1 defines orientations theta in {0,...,5}, and the Limitations section states that tilting is ignored.
  • domain assumption Per-box CoM stability is correctly assessed by projecting the geometric center into the convex hull of support voxels with an alpha margin, with floor boxes unconditionally stable.
    Section 2.3 Center-of-Mass defines the four conditions. The entire training and evaluation pipeline rests on this being a valid physical model.
  • domain assumption The TidyTuesday IKEA scrape and the official IKEA packaging pages provide accurate real-world dimensions and weights.
    Section 1.1 Data Collection uses these sources without independent verification of measurement accuracy.
  • domain assumption A causal language model trained on tokenized coordinates can learn spatial placement from a 5 cm voxel grid.
    Section 2.3 LLP assumes that autoregressive token prediction over discrete coordinates is an adequate representation of the packing policy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PackingGPT: 3D Packing Agent for Real Furniture in Last-Mile Delivery." pith.science (2026). https://pith.science/paper/HE5AVTCQ

@misc{pith2026260801427,
  author       = {Pith},
  title        = {Pith review of: PackingGPT: 3D Packing Agent for Real Furniture in Last-Mile Delivery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HE5AVTCQ}},
  note         = {Machine review of arXiv:2608.01427}
}
read the original abstract

3D bin packing rectangular items into standardised containers to maximise space utilisation under geometric shipping automation. Loading a furniture purchase into a personal vehicle is the same task, but under more complex conditions that standard container loading algorithms ignore. This paper addresses the physically stable placement under these realistic conditions with heterogeneous boxes (e.g. varying dimensions and weights) and occupied containers (e.g. groceries). This paper provides a real-world benchmark dataset and baseline model for the Heterogeneous furniture-in-vehicle packing task. The dataset uses real furniture company flat-pack packaging data covering a large number of catalogue products via family-level extrapolation with diversity length, widths, heights, and weights. We also propose a PackingGPT framework for packing as a sequential placement inspired by the Lego assembly process, where heterogeneous boxes of varying dimensions (bricks) are placed step-by-step into the irregular remaining cargo space (creations). Five baseline packing methods were tested on our dataset without considering the Centre-of- Mass (CoM) constraints. In sedan car simulations, 10-40% of placed boxes failed the stability check on average. When the LLP model was trained on packing sequences with CoM constraints enforced during placement, the failure rate dropped to 0.67% (SUV-500).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

  1. [2003]

    Voice activity detection in the wild via weakly supervised sound event detection

    “On Decomposing 3D Packing Problems in Wooden Furniture Industry.” Proceedings 2003 IEEE International Symposium on Computational Intelligence in Robotics and Automation. Computational Intelligence in Robotics and Automation for the New Millennium (Cat. No.03EX694) 2: 497–502 vol.2. https://doi.org/10.1109/CIRA.2003.1222231. Ma, Ying, Yu Zhou, Qiwei Fang,...

    work page Pith review arXiv 2003

  2. [2013]

    Constraints in Container Loading – a State-of-the-Art Review

    “Constraints in Container Loading – a State-of-the-Art Review.” European Journal of Operational Research 229 (1): 1–20. https://doi.org/https://doi.org/10.1016/j.ejor.2012.12.006. Community, Data Science Learning

    work page doi:10.1016/j.ejor.2012.12.006 2012

  3. [2016]

    A Container Loading Algorithm With Static Mechanical Equilibrium Stability Constraints

    “A Container Loading Algorithm With Static Mechanical Equilibrium Stability Constraints.” Transportation Research Part B: Methodological 91: 565–81. https://doi.org/https://doi.org/10.1016/j.trb.2016.06.003. Yang, Shuo, Shuai Song, Shilei Chu, et al

    work page doi:10.1016/j.trb.2016.06.003 2016

  4. [2023]

    Solving 3D Packing Problem Using Transformer Network and Reinforcement Learning

    “Solving 3D Packing Problem Using Transformer Network and Reinforcement Learning.” Expert Systems with Applications 214: 119153. https://doi.org/https://doi.org/10.1016/j.eswa.2022.119153. Ramos, A. Galrão, José F. Oliveira, José F. Gonçalves, and Manuel P. Lopes

    arXiv 2022

  5. [2024]

    Heuristics Integrated Deep Reinforcement Learning for Online 3D Bin Packing

    “Heuristics Integrated Deep Reinforcement Learning for Online 3D Bin Packing.” IEEE Transactions on Automation Science and Engineering 21 (1): 939–50. https://doi.org/10.1109/TASE.2023.3235742. Zhao, Hang, and Kai Xu

    work page arXiv 2023

  6. [2025]

    A Three-Dimensional Container Loading Algorithm for Solving Logistics Packing Problem

    “A Three-Dimensional Container Loading Algorithm for Solving Logistics Packing Problem.” EURO Journal on Transportation and Logistics 14 (November): 100167. https://doi.org/10.1016/j.ejtl.2025.100167. Pun, Ava, Kangle Deng, Ruixuan Liu, Deva Ramanan, Changliu Liu, and Jun-Yan Zhu

    arXiv 2025

Pith tools

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