Pith. sign in

REVIEW 4 major objections 4 minor 45 references

A legal-action masked DQN plus ZYZ singularity guard lets a 6-DOF co-bot assemble the Soma cube with 75% real-robot success, up from a 35.2% baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A Doosan M0609 cobot learns to assemble a Soma cube with a legal-action-masked DQN and ZYZ singularity guard, achieving 75% real-robot assembly success in 300 trials.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The central 75% real-robot success claim is conditional on the simulator's own success flag, and the paper has enough internal inconsistencies that the evaluation needs major revision before the claimed results can be trusted. the 4 major comments →

arxiv 2508.21272 v1 pith:M5DNIQ6Q submitted 2025-08-29 cs.RO stat.CO

Learning to Assemble the Soma Cube with Legal-Action Masked DQN and Safe ZYZ Regrasp on a Doosan M0609

Jaehong Oh , Seungjun Jung , Sawoong Kim This is my paper
classification cs.RO stat.CO
keywords Soma cube assemblymasked deep Q-networklegal-action maskingZYZ singularity avoidanceregrasp planningcurriculum learningsim-to-real transfercollaborative robot manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that a 6-DOF collaborative robot can learn to assemble the seven-piece Soma cube entirely from reinforcement learning, if the learning is constrained by a legal-action mask and protected by singularity-aware regrasping. On a real Doosan M0609 with an underactuated gripper, the integrated system reports 75.0% end-to-end assembly success over 300 trials (95% CI [70.1, 79.9]), up from a 35.2% baseline, with average completion time 12.3 minutes and ±1.8 mm placement accuracy. The central technical claim is that decomposing the Q-function into orientation and position components, plus masking out physically illegal actions, reduces the combinatorial action space from 3,132 to 2,484 while preserving all physically executable solutions. A ZYZ proximity-index guard is claimed to raise motion success from 54% to 96% by triggering a six-step regrasp before gimbal-lock configurations arise. If these results hold, they provide a practical template for constraint-aware RL in real robotic assembly, where the bottleneck is usually not learning but the gap between abstract actions and feasible motions.

Core claim

On the paper's own terms, the discovery is that physical constraints can be turned from obstacles into structure: a legal-action mask makes the DQN's action space smaller and safer without cutting off optimal solutions, and a proximity index on ZYZ Euler angles converts singular orientations into planned regrasps instead of crashes. The combined system—masked hierarchical DQN, curriculum from 2 to 7 pieces, dense robot-friendly rewards, global mapping, and speech commands—achieves 75% success on the real robot, with a sim-to-real transfer efficiency of 91%. The action-space reduction from O(3,132) to O(116)+O(27) parameters and the 26% improvement in sample efficiency are presented as the co

What carries the argument

The load-bearing mechanism is the legal-action mask combined with hierarchical Q decomposition. The mask m(s) is a conjunction of four constraint predicates—collision, support, reachability, and vertical access—that zeroes out infeasible actions before the argmax; Q is factored as Q(s,(o,p)) = Q_ori(s,o) + Q_pos(s,p), cutting the output layer from 3,132 to 143. The second mechanism is the ZYZ singularity guard: a proximity index PI(β)=1−|cos β| detects near-singular configurations and triggers a six-step regrasp (wrist clearance, retraction, corrective rotation, pose alignment, fallback) so that Euler-angle motion plans remain executable.

Load-bearing premise

The reported 75% real-robot success rate is measured only on episodes the simulator already flagged as successful (Section VI.A.1: real-robot invocation is selective, with 5,247 policy calls from 100,000 simulation episodes), so the headline number is conditional on the simulator's own judgment and may not represent the policy's success on uncurated inputs.

What would settle it

Run 300 real-robot trials where the simulator's success flag is ignored and every episode is executed regardless of the simulator's prediction; if the success rate falls below the claimed 95% CI [70.1, 79.9] or near the 35.2% baseline, the headline claim is not established. As a secondary check, count how many of the reported 300 trials were drawn from the 5,247 selective invocations; if fewer than 300 independent, uncurated episodes were physically executed, the claimed statistical precision is unsupported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A masked, factored DQN can be trained to full Soma-cube assembly in simulation with only 39.9% Level-3 success, yet still transfer to a real robot at 75% end-to-end success, suggesting that imperfect policies can be made deployable with sufficient safety planning.
  • The legal-action mask's 26% sample-efficiency gain is directly transferable to other discrete manipulation tasks with combinatorial action spaces, such as modular or circuit-board assembly.
  • The ZYZ proximity-index guard with six-step regrasp offers a reusable recipe for any 6-DOF Euler-angle controller to avoid gimbal-lock failures, improving motion success from 54% to 96%.
  • The curriculum sequence (2-piece, 3-piece, 7-piece) with a ground-first, vertically accessible reward shape is a training schedule that other multi-piece assembly learners could adopt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported 75% should be read as conditional on the simulator's success flag, because real-robot invocation was selective; an uncurated end-to-end test on random inputs would likely show a lower rate, and such a test is needed before generalizing the headline.
  • The additive Q decomposition assumes orientation and position preferences are independent given the state; in more tightly coupled assemblies this assumption may break, so a coupling-aware factorization is the natural next extension.
  • The tri-modal reward distribution suggests the policy has multiple solution strategies; exploiting this diversity by switching strategies when one path fails could push the success rate beyond the current ceiling.
  • The legal-action mask could be generalized from a hard binary filter to a learned soft constraint, which would let the robot trade off feasibility against reward in continuous or hybrid action spaces.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents an integrated Soma-cube assembly system built around a legal-action-masked DQN, a ZYZ singularity guard, Unity-based global mapping, and Whisper speech recognition, deployed on a Doosan M0609 robot. The main claims are a 75.0% end-to-end real-robot assembly success rate over 300 trials (95% CI [70.1, 79.9]), a reduction of the action space from 4,536/3,132 actions to about 2,484 while "maintaining solution completeness," a 26% sample-efficiency improvement, and a 91% sim-to-real transfer efficiency. The paper also describes curriculum learning, reward shaping, an ablation study, and component-level evaluations of vision, motion planning, and speech recognition.

Significance. If the central empirical claim were valid, the work would be a useful engineering demonstration of an integrated learned manipulation pipeline: constraint-based action masking, singularity-aware regrasp, real-time perception, and speech-based HRI on a real cobot. The component-level data (YOLO mAP, ZYZ regrasp improvement, speech accuracy, and the ablation table) are potentially informative, and the appendix provides unusually detailed hyperparameters and protocols. However, the headline result is not established as stated: the physical robot is invoked only after the simulator has flagged an episode as successful, so the reported 75% is at best a conditional success rate. The paper also contains serious internal inconsistencies in its failure statistics, reward functions, action-space sizes, and state-space dimensions. These problems affect the paper's central claims and make the current manuscript unsuitable for publication without substantial re-evaluation.

major comments (4)
  1. [VI.A.1 and VI.G.2] The headline "end-to-end assembly success rate of 75.0% over 300 trials" (Eq. 40) is not supported by the described protocol. Section VI.A.1 says the real robot is invoked selectively: "when the RL agent's success flag transitions to True ... the system executes the planned sequence on the physical robot," yielding 5,247 robot policy calls from 100,000 simulation episodes. If the 300 trials are drawn from this filtered set, the reported proportion estimates P(robot success | simulator success flag), not P(robot success) from randomized initial states. The paper never reports the denominator: total random-start episodes, number flagged successful, number physically executed, and successes at each stage. Without an uncurated evaluation or an explicit decomposition, the central empirical claim is not established.
  2. [Table I] The failure-mode table is internally inconsistent. The Frequency column sums to 165% (12+16+10+26+6+4+16+75), while the Count column sums to 300 with success count 225. If the percentages are percentages of all 300 trials, they do not match the counts (e.g., 12% would be 36, not 9). If the failure percentages are percentages of the 75 failures, they should sum to 100 and the counts imply 12.0/16.0/10.7/26.7/5.3/4.0/25.3, not the listed values. The table therefore does not support the failure-mode analysis and the statistics need to be recomputed and clearly defined.
  3. [IV.A.2.a and IV.A.2.b] The "Solution Completeness Guarantee" is definitional rather than a substantive guarantee. Legal actions are defined in Eq. (5)/(17) as actions satisfying exactly the physical constraints (no collision, support, reachability, vertical access). The claim that any "physically executable" solution uses only legal actions is then either tautological, if "physically executable" is defined to be the same conjunction, or unproven, if vertical access is added as an extra restriction. Similarly, Theorem 1 states the exploration-efficiency improvement as the set-size ratio |A_total|/|A_legal| ≈ 1.26; this is a static counting ratio, not a bound on sample efficiency. The actual convergence benefit should be established from learning curves or ablations; Table III reports a 46% reduction in episodes to converge, which is inconsistent with the 26% figure. Please provide an independent definition of c
  4. [III.B, IV.A.1, VI.C.1] Several core definitions are inconsistent across the paper. The action space is 3,132 in Section III.B (116 orientations × 27 positions), but Section VI.C.1 says "7×24×27 = 4,536 combinations" and the Introduction says 4,536 theoretical actions. The state dimension is 36 in Eq. (9) but 34 in Section VI.C.1. Three different reward functions are given: Eq. (6), Eq. (22), and Eq. (41). These inconsistencies affect the claimed 26% sample-efficiency improvement and make the method difficult to reproduce. Please reconcile all definitions and recompute the complexity and sample-efficiency numbers.
minor comments (4)
  1. [VI.A.2] The success criterion "Temporal performance requires completion within 100ms control cycles" is ambiguous: it appears to mix a control-cycle latency requirement with the assembly completion time target of 15 minutes stated in Section III.D. Please clarify.
  2. [VII.A.1 and VI.A.1] The total number of training episodes is given as 100,000 in Section VI.A.1 but 105,300 in the abstract and Section VII.A.1; the curriculum-level counts (500 + 1,600 + 102,100 = 104,200) are inconsistent as well.
  3. [Appendix A] The "Code Repository and Data Availability" section describes a repository structure but provides no repository URL or access instructions. Please include the actual link or state that code is available on request.
  4. [Figures] Figures 3 and 5–11 are referenced in the text, but the manuscript as submitted does not appear to include the actual rendered figures; please ensure all figures are present and legible.

Circularity Check

3 steps flagged

Headline 75% real-robot success is conditional on the simulator's own success flag; the solution-completeness guarantee is true by definition; and the claimed 26% sample-efficiency gain is just the ratio of action-space sizes.

specific steps
  1. self definitional [Section IV.A.2.a (Solution Completeness Guarantee; see Eqs. 5 and 17)]
    "The legal action space Alegal(s) = {a : m(s)[a] = 1 } maintains solution completeness under the constraint that physical manipulator limitations are preserved. Formally, for any valid Soma cube solution π∗ = ( s0, a∗0, s1, a∗1, ..., s∗T ) that is physically executable by the robot, we guarantee a∗t ∈ Alegal(st) for all t ∈ [0, T]"

    Legal actions are defined by a conjunction of physical constraints: m(s)[a] = ∧ c(s,a) with C = {collision, support, reachability, vertical access} (Eq. 17; Eq. 5). A solution that is 'physically executable' is exactly one whose actions satisfy those constraints. Therefore the promised containment in Alegal is true by definition; the guarantee adds no independent content and cannot fail. It is a restatement of the definition of legal action, not a derived completeness theorem.

  2. self definitional [Section IV.A.2.b (Theorem 1, Sample Efficiency Analysis)]
    "Theorem 1. Legal action masking improves exploration efficiency by approximately |Atotal|/|Alegal| where |Atotal| = 3, 132 is the complete action space and |Alegal| ≈2, 484 is the average number of feasible actions per state. Proof. ... The exploration efficiency improvement ratio is 3,132/2,484 ≈ 1.26, representing a 26% reduction in required exploration samples..."

    The claimed 26% sample-efficiency improvement is not derived from any sample-complexity bound or independent training measurement; it is literally the ratio of the two action-space cardinalities stated as the theorem's premises. The 'proof' rewrites the action-space reduction as 'exploration efficiency.' Unless sample efficiency is defined as this ratio (which would make the claim definitional), no exploration-efficiency conclusion follows.

  3. other [Sections VI.A.1 and VI.G.2 (Real robot invocation protocol and End-to-end assembly success rate)]
    "Real robot invocation occurs selectively based on internal policy evaluation: when the RL agent’s success flag transitions to True ... the system executes the planned sequence on the physical robot. This selective deployment protocol resulted in 5,247 actual robot policy calls from 100,000 total simulation episodes... End-to-end assembly evaluation over 300 independent trials achieves 75.0% ±4.9% success rate"

    The reported 75% is measured only on episodes that the RL agent's own success flag has already labeled as successful in simulation. Consequently the metric is P(robot success | simulator success flag), not P(success) over the 300 randomized initial-state trials claimed. The denominator of total random-start episodes, number flagged, number physically executed, and successes at each stage is never given. The headline 'end-to-end' success rate is therefore conditional on the system's own success predicate—a self-selected evaluation set—rather than an independent measure of system performance.

full rationale

The paper's core engineering contributions (masked DQN training, ZYZ singularity guard, Unity mapping, Whisper HRI) are empirical and not circular; they are validated by measurements that could in principle falsify them. No self-citation chain is involved. However, three load-bearing claims fail the circularity test. (1) The 'Solution Completeness Guarantee' is a definitional tautology: legal actions are defined as actions satisfying collision/support/reachability/vertical-access constraints, so any solution that is physically executable is legal by construction. (2) 'Theorem 1' does not prove a sample-efficiency bound; it states the ratio of action-space sizes and calls the ratio an exploration-efficiency improvement, so the 26% is just the action-space reduction renamed. (3) The headline 75% real-robot success rate is conditional on the simulator's own success flag because real-robot invocation is gated on that flag; the paper never reports the denominator of random-start episodes, flagged episodes, and physically executed episodes. The reported CI [70.1, 79.9] therefore characterizes a self-selected subset, not the claimed end-to-end system performance. These are partial circularities; the central empirical pipeline still has independent content, so score 6.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central empirical claims rest on hand-tuned reward coefficients, hand-set singularity thresholds, and unproven assumptions about the Q-function factorization and the completeness of the mask. No independent evidence or sensitivity analysis is provided.

free parameters (3)
  • Reward shaping coefficients = Rbase=10, Rground=30/25/0, Raccess=8/-30, Rheight=-8, Rlogic=±15, Rstructure=2
    Hand-tuned values in Eqs. 23-28 encode the ground-first assembly policy; no sensitivity analysis or learning-based tuning is reported.
  • Proximity index threshold = PI>0.9, |β|>βthreshold, clamp to ±89.9°
    Arbitrary threshold in Algorithm 3 to trigger regrasp; chosen by hand, no robustness study.
  • Simplified reward coefficients (Eq. 41) = +100, +1, -5 to -10, 0
    Second reward specification differs from Eq. 6 and Eqs. 22-28; conflicting definitions of the reward.
axioms (4)
  • ad hoc to paper Q-function factorizes additively over orientation and position (Eq. 15)
    Section IV.A.1.a asserts orientation and position preferences are approximately independent given the state; this is not proven and is questionable for geometric assembly where orientation determines occupied cells.
  • ad hoc to paper Legal action mask preserves all physically executable solutions (Solution Completeness Guarantee)
    Section IV.A.2.a asserts the guarantee without proof; it follows by definition of the mask as the set of feasible actions.
  • domain assumption Simulation-to-real transfer fidelity via domain randomization
    Section V assumes PyBullet modeling and DR parameters yield a 91% transfer efficiency; no closed-loop sim-to-real analysis is provided.
  • domain assumption ZYZ singularity handling via clamping to ±89.9° and 6-step regrasp is safe and effective
    Algorithms 2-3 use hand-set thresholds; no formal safety proof.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Assemble the Soma Cube with Legal-Action Masked DQN and Safe ZYZ Regrasp on a Doosan M0609." pith.science (2026). https://pith.science/paper/M5DNIQ6Q

@misc{pith2026250821272,
  author       = {Pith},
  title        = {Pith review of: Learning to Assemble the Soma Cube with Legal-Action Masked DQN and Safe ZYZ Regrasp on a Doosan M0609},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5DNIQ6Q}},
  note         = {Machine review of arXiv:2508.21272}
}
Share X Bluesky LinkedIn Reddit HN
abstract

This paper presents the first comprehensive application of legal-action masked Deep Q-Networks with safe ZYZ regrasp strategies to an underactuated gripper-equipped 6-DOF collaborative robot for autonomous Soma cube assembly learning. Our approach represents the first systematic integration of constraint-aware reinforcement learning with singularity-safe motion planning on a Doosan M0609 collaborative robot. We address critical challenges in robotic manipulation: combinatorial action space explosion, unsafe motion planning, and systematic assembly strategy learning. Our system integrates a legal-action masked DQN with hierarchical architecture that decomposes Q-function estimation into orientation and position components, reducing computational complexity from $O(3,132)$ to $O(116) + O(27)$ while maintaining solution completeness. The robot-friendly reward function encourages ground-first, vertically accessible assembly sequences aligned with manipulation constraints. Curriculum learning across three progressive difficulty levels (2-piece, 3-piece, 7-piece) achieves remarkable training efficiency: 100\% success rate for Level 1 within 500 episodes, 92.9\% for Level 2, and 39.9\% for Level 3 over 105,300 total training episodes.

Figures

Figures reproduced from arXiv: 2508.21272 by Jaehong Oh, Sawoong Kim, Seungjun Jung.

Figure 1
Figure 1. Figure 1: The seven distinct Soma cube pieces with their geometric properties [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Real-world deployment of our RL-based Soma cube assembly [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Legal action masking visualization showing constraint-based action [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Complete system architecture showing the integration of legal-action [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reinforcement learning training visualization showing the progres [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training progression showing success rate evolution across 105,300 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scatter plot showing positive correlation (r=0.495) between episode [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: DQN training loss evolution showing convergence with beneficial [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Epsilon-greedy exploration decay schedule transitioning from explo [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Gaussian Mixture Model analysis revealing three distinct reward [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Collaborative robot market size and growth analysis 2024-2033,

    Market Research Intellect, “Collaborative robot market size and growth analysis 2024-2033,” Market Research Report , 2024

    work page 2024

  2. [2]

    Introduction to robotics: mechanics and control,

    J. J. Craig, “Introduction to robotics: mechanics and control,” Pearson Education, 2005

    work page 2005

  3. [3]

    Modern robotics: mechanics, planning, and control,

    K. M. Lynch and F. C. Park, “Modern robotics: mechanics, planning, and control,” Cambridge University Press , 2017

    work page 2017

  4. [4]

    Reinforcement learning: An introduction,

    R. S. Sutton and A. G. Barto, “Reinforcement learning: An introduction,” MIT press, 2018

    work page 2018

  5. [5]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V . Panneershelvam, M. Lanctot et al., “Mastering the game of go with deep neural networks and tree search,” in nature, vol. 529, no. 7587. Nature Publishing Group, 2016, pp. 484–489

    work page 2016

  6. [6]

    Grand- master level in starcraft ii using multi-agent reinforcement learning,

    O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi, R. Powell, T. Ewalds, P. Georgiev et al., “Grand- master level in starcraft ii using multi-agent reinforcement learning,” Nature, vol. 575, no. 7782, pp. 350–354, 2019

    2019

  7. [7]

    Outracing champion gran turismo drivers with deep reinforcement learning,

    P. R. Wurman, S. Barrett, K. Kawamoto, J. MacGlashan, K. Subrama- nian, T. J. Walsh, R. Capobianco, A. Devlic, F. Eckert, F. Fuchs et al., “Outracing champion gran turismo drivers with deep reinforcement learning,” vol. 602, no. 7896. Nature Publishing Group, 2022, pp. 223–228

    work page 2022

  8. [8]

    End-to-end training of deep visuomotor policies,

    S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of deep visuomotor policies,” in The Journal of Machine Learning Research, vol. 17, no. 1, 2016, pp. 1334–1373

    work page 2016

  9. [9]

    Scalable deep reinforcement learning for vision-based robotic manipulation,

    D. Kalashnikov, A. Irpan, P. Pastor, J. Ibarz, A. Herzog, E. Jang, D. Quillen, E. Holly, M. Kalakrishnan, V . Vanhoucke et al. , “Scalable deep reinforcement learning for vision-based robotic manipulation,” in Conference on robot learning . PMLR, 2018, pp. 651–673

    work page 2018

  10. [10]

    Robotics: mod- elling, planning and control,

    B. Siciliano, L. Sciavicco, L. Villani, and G. Oriolo, “Robotics: mod- elling, planning and control,” Springer Science & Business Media, 2009

    work page 2009

  11. [11]

    Probabilistic robotics,

    S. Thrun, W. Burgard, and D. Fox, “Probabilistic robotics,” MIT press, 2005

    work page 2005

  12. [12]

    Transporter networks: Rearranging the visual world for robotic manipulation,

    A. Zeng, P. Florence, J. Tompson, S. Welker, J. Chien, M. Attarian, T. Armstrong, I. Krasin, D. Duong, V . Sindhwani et al. , “Transporter networks: Rearranging the visual world for robotic manipulation,” in Conference on Robot Learning . PMLR, 2020, pp. 726–747

    work page 2020

  13. [13]

    Near-optimal represen- tation learning for hierarchical reinforcement learning,

    O. Nachum, S. Gu, H. Lee, and S. Levine, “Near-optimal represen- tation learning for hierarchical reinforcement learning,” arXiv preprint arXiv:1810.01257, 2018

    Pith/arXiv arXiv 2018

  14. [14]

    Masked reinforcement learning with action masking,

    S. Huang, R. F. J. Dossa, C. Ye, J. Braga, D. Chakraborty, K. Mehta, and J. G. M. Araújo, “Masked reinforcement learning with action masking,” arXiv preprint arXiv:2006.14171 , 2022

    Pith/arXiv arXiv 2006

  15. [15]

    Efficient algo- rithms for polyhedral assembly partitioning,

    M. Bertram, G. Reis, R. van Liere, and P. Slusallek, “Efficient algo- rithms for polyhedral assembly partitioning,”Computer Graphics Forum, vol. 27, no. 2, pp. 235–244, 2008

    work page 2008

  16. [16]

    Redshifts versus paradigm shifts; against renaming Hubble's Law

    Y . Li, J. Wu, R. Tedrake, J. B. Tenenbaum, and A. Torralba, “Learning to assemble objects with robot hands,” arXiv preprint arXiv:1909.07731, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  17. [17]

    Computer recreations: solid geometry, billiard-ball computers and other things,

    A. K. Dewdney, “Computer recreations: solid geometry, billiard-ball computers and other things,” Scientific American , vol. 252, no. 6, pp. 18–27, 1985

    work page 1985

  18. [18]

    Prob- abilistic roadmaps for path planning in high-dimensional configuration spaces,

    L. E. Kavraki, P. Švestka, J.-C. Latombe, and M. H. Overmars, “Prob- abilistic roadmaps for path planning in high-dimensional configuration spaces,” in Proceedings of the 1996 IEEE International Conference on Robotics and Automation , vol. 2. IEEE, 1996, pp. 566–580

    work page 1996

  19. [19]

    Find- ing locally optimal, collision-free trajectories with sequential convex optimization,

    J. Schulman, J. Ho, A. Lee, I. Awwal, H. Bradlow, and P. Abbeel, “Find- ing locally optimal, collision-free trajectories with sequential convex optimization,” Robotics: science and systems , pp. 1–10, 2013

    work page 2013

  20. [20]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” in International Conference on Machine Learning . PMLR, 2023, pp. 28 492–28 518

    work page 2023

  21. [21]

    Using natural language for personalized human-robot interaction,

    A. L. Thomaz, G. Hoffman, and M. Cakmak, “Using natural language for personalized human-robot interaction,” KI-Künstliche Intelligenz , vol. 30, no. 2, pp. 121–130, 2016

    work page 2016

  22. [22]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  23. [23]

    Deep reinforcement learning with double q-learning,

    H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double q-learning,” Proceedings of the AAAI conference on artifi- cial intelligence, vol. 30, no. 1, 2016

    work page 2016

  24. [24]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” Proceedings of the 26th annual international conference on machine learning, pp. 41–48, 2009

    work page 2009

  25. [25]

    Reverse curriculum generation for reinforcement learning,

    C. Florensa, D. Held, M. Wulfmeier, M. Zhang, and P. Abbeel, “Reverse curriculum generation for reinforcement learning,” Conference on robot learning, pp. 482–495, 2017

    work page 2017

  26. [26]

    Reinforcement learning in robotics: A survey,

    J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” in The International Journal of Robotics Research , vol. 32, no. 11. Sage Publications Sage UK: London, England, 2013, pp. 1238–1274

    work page 2013

  27. [27]

    Dueling network architectures for deep reinforcement learning,

    Z. Wang, T. Schaul, M. Hessel, H. Hasselt, M. Lanctot, and N. Fre- itas, “Dueling network architectures for deep reinforcement learning,” International conference on machine learning , pp. 1995–2003, 2016

    work page 1995

  28. [28]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  29. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 770–778, 2016

    work page 2016

  30. [30]

    Dropout: a simple way to prevent neural networks from over- fitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: a simple way to prevent neural networks from over- fitting,” The journal of machine learning research , vol. 15, no. 1, pp. 1929–1958, 2014

    work page 1929

  31. [31]

    Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates,

    S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates,” 2017 IEEE international conference on robotics and automation (ICRA), pp. 3389–3396, 2017

    work page 2017

  32. [32]

    Learn- ing dexterous in-hand manipulation,

    M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray et al., “Learn- ing dexterous in-hand manipulation,” in The International Journal of Robotics Research, vol. 39, no. 1, 2020, pp. 3–20. APPENDIX A IMPLEMENTATION DETAILS AND REPRODUCIBILITY This appendix provides comprehensive implementa...

    work page 2020

  33. [33]

    Hardware Requirements: The complete system operates on the following hardware configuration: • Robot Platform : Doosan M0609 (6-DOF, 6kg payload, 900mm reach, ±0.05 mm repeatability) • End Effector: OnRobot RG2 2F Gripper (110mm stroke, 40N max force, 0.1N force resolution) • Vision System : Intel RealSense D435i (RGB: 1920 × 1080 @30fps, Depth: 1280 × 72...

    work page 1920

  34. [34]

    Detailed Hyperparameter Configuration

    Software Dependencies: The software stack requires the following specific versions for reproducibility: • Operating System : Ubuntu 22.04 LTS with RT kernel patches • ROS Framework: ROS2 Humble Hawksbill (DDS: Cy- cloneDX, QoS: RELIABLE/BEST_EFFORT) • Robot Control : Doosan Robot SDK v2.1, MoveIt2 v2.5.4 • Machine Learning : PyTorch 1.13.1, CUDA 11.8, cuD...

    work page 2022

  35. [35]

    DQN Training Parameters: Table IV presents the com- plete hyperparameter configuration used for DQN training: TABLE IV DQN H YPERPARAMETER CONFIGURATION Parameter Value Learning Rate ( α) 1 × 10−4 Discount Factor ( γ) 0.99 Initial Epsilon ( εstart) 0.9 Final Epsilon ( εend) 0.1 Epsilon Decay Steps 40,000 Target Network Update ( τ) 20 episodes Replay Buffe...

    work page

  36. [36]

    Reward Function Coefficients: The reward function em- ploys the following carefully tuned coefficients: rcomplete = +100 (puzzle completion) (45) rdensity = +1 (valid placement increasing density) (46) rinvalid = −8 (invalid action baseline) (47) rcollision = −10 (collision violation) (48) rboundary = −5 (boundary violation) (49) rgravity = −12 (gravity v...

    work page

  37. [37]

    Root Cause Analysis and Solution Architecture

    Vision System Parameters: Critical vision processing parameters include: • YOLOv8n: Input resolution 640×640, confidence thresh- old 0.5, NMS threshold 0.4 • Camera Calibration: 12 chessboard poses, 9×6 pattern, 25 mm squares • Hand-Eye Transform : Tsai-Lenz method, RMS error target <2mm • Depth Processing : Temporal filter (5 frames), spatial filter (3 ×...

    work page

  38. [38]

    Failure Analysis Framework: The systematic failure analysis revealed four primary root causes (RCA): RCA-1: Combinatorial Action Space Explosion • Problem: Single Q(s,a) network handling 4,536 theoret- ical actions • Solution: Action decomposition into Orientation-Q + Position-Q with summation argmax • Impact: 40% reduction in exploration time, 26% im- pr...

    work page

  39. [39]

    Experimental Protocols and Quality Assurance

    Improved System Architecture: The final architecture implements a three-stage pipeline: Environment → Policy → Execution, with the following key components: • Environment Stage : Legal action masking, constraint validation, state normalization • Policy Stage: Decomposed Q-networks (Q_ori + Q_pos), experience replay with PER • Execution Stage : Rotation sp...

    work page

  40. [40]

    Training Protocol: The standardized training protocol ensures reproducible results: Algorithm 4 Reproducible DQN Training Protocol Require: Random seeds: numpy.seed(42), torch.manual_seed(42), random.seed(42) 1: Initialize environment with constraint masking 2: Initialize DQN with specified architecture 3: Initialize replay buffer (capacity: 50,000) 4: Lo...

    work page

  41. [41]

    Evaluation Metrics and Validation: Comprehensive evaluation employs multiple metrics: • Success Rate: Binary completion of 3×3×3 Soma Cube (target: >70%) • Assembly Time : Total time from start command to completion (target: <15 minutes) • Position Accuracy : Final piece placement error (target: ±2 mm) • Grasp Success Rate : Successful piece manipulation ...

    work page

  42. [42]

    Code Repository and Data Availability

    Quality Assurance Checklist: The following checklist ensures system reliability: □ All 27 cube positions verified for geometric consistency □ Legal action mask unit tests pass for bound- ary/overlap/support constraints □ Learning curves reproducible within ±3% across 5 inde- pendent runs □ Coordinate transformation error <2mm threshold main- tained □ ZYZ ...

    work page

  43. [43]

    Repository Structure: The complete implementation is organized with the following directory structure: Source Code Organization: • src/dqn_agent/: RL training and inference modules • src/vision_pipeline/: YOLO detection and pose estimation • src/motion_planning/: ZYZ regrasp and path planning • src/robot_control/: Doosan SDK interface • src/unity_mapping/...

    work page

  44. [44]

    The setup process includes automated de- pendency installation, robot calibration procedures, and system validation tests

    Installation and Setup: Detailed installation instruc- tions are provided in the repository documentation, includ- ing Docker containerization for consistent deployment across different systems. The setup process includes automated de- pendency installation, robot calibration procedures, and system validation tests

    work page

  45. [45]

    Performance Benchmarks: Reference performance benchmarks are provided for system validation: • Training Time: 72 hours on RTX 4090 for 50k episodes • Inference Speed: 12ms DQN forward pass, 23ms YOLO detection • Memory Usage: Peak 3.6GB during concurrent training and visualization • Network Bandwidth : 80MB/s for ROS-Unity point cloud streaming This compr...

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.