Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A dual-arm robot can now assemble general multi-part objects from CAD models alone.

desk verdict A substantial dual-arm assembly system with a genuinely useful benchmark, but the abstract claims complete autonomous assembly while the paper's own Table 5 shows only 15% zero-intervention success — the claims need to be pulled back to match the data. read the letter →

arxiv 2506.05168 v1 pith:M4MUSVHP submitted 2025-06-05 cs.RO

Yunsheng Tian , Joshua Jacob , Yijiang Huang , Jialiang Zhao , Edward Gu , Pingchuan Ma , Annan Zhang , Farhad Javid
show 5 more authors
Branden Romero Sachin Chitta Shinjiro Sueda Hui Li Wojciech Matusik
This is my paper · ORCID
classification cs.RO
keywords multi-partassemblydual-armmanipulationplanningreinforcementlearningsim-to-realtransferequivariantpolicyfixturegenerationbenchmark
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

Fabrica claims to be the first robotic system that maps a CAD model of a 5-to-9-part object to a physical assembly performed by two arms, with no domain knowledge or human demonstrations. It splits the problem: a hierarchical planner decides the order, grasps, holding roles, fixtures, and motions, while a single reinforcement-learned policy handles the contact-rich insertion steps. The policy transfers zero-shot to the real robot and succeeds on 80% of individual assembly steps; across whole assemblies, two human interventions raise completion from near zero to 95%. If the claim holds, general-purpose assembly—not just two-part top-down insertion—becomes tractable for dual-arm robots.

What carries the argument

The load-bearing mechanism is the path-centric coordinate transformation, an SE(3)-equivariant remapping of each planned insertion path into a canonical top-down frame; combined with residual actions derived from the plan, it lets a single policy learn one insertion skill that transfers across geometries, directions, and grasps. The hierarchical planner supplies stable sequences and feasible grasps that keep errors small enough for the policy to correct, and the fixture generator ensures parts are picked up without regrasping.

What would settle it

Measure the actual distribution of part-pose errors at the start of each insertion on the physical setup, and count insertions that fail specifically because the holding gripper blocks the path. If the median error exceeds the 3 mm training envelope or obstruction failures do not decrease with retries, the zero-shot transfer claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that complete, generalizable multi-part assembly can be achieved by integrating an efficient global planner with local equivariant reinforcement learning policies, rather than by engineering a single monolithic controller. The planner optimizes a discrete-continuous objective over assembly-hold sequences, grasps, and motions, ensuring that every insertion has a stable support part and a feasible dual-arm configuration. Contact-rich insertions are handled by a generalist policy trained in simulation, using a path-centric coordinate transformation that reorients any straight-line insertion into a canonical top-down task, and residual actions that add corrective motion to the plan. In real-world tests on seven objects, the system reaches 80% step-level success and 95% multi-step completion with two interventions; an out-of-distribution version of the policy, trained on other assemblies, matches this performance, which the authors take as evidence of generalization.

Load-bearing premise

The system's real-world success rests on the assumption that 3 mm of simulated pose noise plus retries covers the actual positioning errors and the unmodeled holding gripper's obstruction; if real errors are larger, the learned policy cannot correct them and whole assemblies fail without human help.

Editorial extensions

If this is right

  • A new CAD model plus robot setup can be turned into a complete assembly program in minutes, without manual programming or demonstrations.
  • One generalist policy trained across all seven benchmark assemblies matches the performance of per-assembly specialists, so insertion skills transfer across object geometries and directions.
  • An out-of-distribution policy trained on six assemblies performs as well on the seventh, suggesting a pretrained library of insertion skills can be reused for novel objects.
  • Optimizing part sequences and grasps materially improves control stability and real-world success, as shown by ablations that remove either optimization.
  • Allowing up to three attempts per insertion step substantially raises step success, meaning the policies benefit from retry rather than requiring single-trial perfection.

Reading between the lines

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

  • If the 3 mm randomization assumption holds, the same architecture should extend to bin-picking by replacing the fixture with a vision system: the path-centric policy is agnostic to how the part is initially placed, so failures would concentrate in perception, not insertion.
  • The path-centric residual formulation is a general recipe: any contact-rich skill with a known straight-line approach path, such as screwing or sliding, could reuse the same policy representation, though the paper only tests insertion.
  • The near-zero no-intervention end-to-end rate suggests the bottleneck is error accumulation and the unmodeled holding gripper, not single-step skill; adding the holding gripper to simulation or real-time perception would be the natural next experiment.
  • Because generalist policies trained out-of-distribution matched specialists, the paper implies a library of insertion skills can be pretrained once and reused for new objects, but a road-test would be assembling a never-seen 10+ part object without any retraining.
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 presents Fabrica, a dual-arm robotic system that maps CAD assembly models to physical execution through a hierarchical planner (precedence, grasp, sequence, fixture, and motion planning) and learned insertion policies. The RL component uses path-centric coordinate transformations, residual actions, and a minimal distance-based reward, with policies trained in simulation and deployed zero-shot in the real world. The paper introduces a seven-assembly benchmark and reports step-level success, multi-step cumulative success with zero/one/two human interventions, and ablation studies isolating the contributions of sequence optimization, grasp optimization, path-centric transforms, and residual actions.

Significance. If the stated results are taken at face value, Fabrica is a useful integration of planning and learning: the automated fixture design, the generalist policy trained across assemblies and grasp poses, the leave-one-assembly-out OOD evaluation, and the explicit failure analysis are all valuable contributions. The planner's cost functions f1-f4 and the policy's reward are clearly defined, and the ablations are informative. However, the headline claim of complete autonomous assembly is not supported by the paper's own numbers: zero-intervention end-to-end success is 15% overall for the AS policy and 10% for the AG policy, and the paper states that all methods achieve near-zero multi-step success without intervention. The contribution is real but the claims need substantial qualification before the paper can be accepted.

major comments (3)
  1. [Abstract, Sec. 5.4, Table 5] The abstract states that Fabrica is 'the first system to achieve complete and generalizable real-world multi-part assembly without domain knowledge or human demonstrations,' but Table 5 reports zero-intervention cumulative success of 15% for AS and 10% for AG, and Sec. 5.4 states that 'all methods achieve near-zero multi-step success rates without intervention.' The 95%/100% two-intervention numbers require human rescue after three failed insertion attempts, with only three runs per assembly, so many table entries correspond to a single successful run (e.g., 33%). The 'complete assembly' claim is therefore not supported by the reported data as stated; the abstract and introduction should be revised to report the zero-intervention results and to qualify the 'complete' claim as requiring up to two human interventions.
  2. [Abstract, Table 4, App. E.2] The abstract's '80% successful steps' figure is the Table 4 overall number obtained with up to three trials per step and state-based success detection, not with a single policy execution. Appendix E.2 shows that with one trial, the AS policy achieves only 58%, 58%, 60%, 40%, 67%, 43%, and 75% step success on the seven assemblies (average about 57%), a substantial difference. The paper should either report single-trial step success as the headline or explicitly state that retries are part of the deployment protocol.
  3. [Sec. 4.3, Sec. 5.4, App. D.4] The zero-shot transfer claim is weakened by the paper's own failure analysis. Policies are trained with 3 mm pose noise (Sec. 4.3), but Sec. 5.4 reports that 'many sources of real-world error lead to much more significant errors than simulated' and that the unmodeled holding gripper 'causes unexpected part obstructions in the real world.' These statements, together with the near-zero zero-intervention completion, indicate that the 3 mm randomization does not by itself cover the real deployment distribution. To support the zero-shot transfer claim, the paper should provide a measured distribution of real pre-insertion pose errors and a quantitative comparison with the simulated noise level, or soften the claim accordingly.
minor comments (5)
  1. [Sec. 5.4, Tables 4-5] The text says 'we allow up to three trials per step until success' while the table captions say 'without intervention'; please use 'human intervention' for human resets and clarify that policy retries are not counted as interventions.
  2. [Sec. 3, App. C.3] The paper claims 'optimality guarantees under assumptions (A1)-(A5)' for the hierarchical planning scheme, but no formal theorem or proof is given; either provide a precise optimality statement with a proof or soften the claim to 'optimal among the enumerated feasible sequences under the stated assumptions.'
  3. [Appendix F] The VLM-based corrective alignment results are presented with qualitative examples but are not integrated into the main Tables 4-5; please state explicitly that these results are preliminary and are not part of the reported benchmark numbers.
  4. [Abstract, Sec. 1] The claim that the system is general to 'any dual-arm robots' is stronger than the evidence: real experiments use one Panda setup and other robots are demonstrated only in simulation; please qualify this statement.
  5. [Appendix F caption] There is a typo in the caption 'Physical setup for integerat-ing vision feedback'; it should read 'integrating.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the planning and learning pipeline is evaluated against external tasks and metrics, with no fitted parameters renamed as predictions or load-bearing self-citation chains.

full rationale

Fabrica's derivation chain is self-contained against external benchmarks. The planner minimizes explicit, a priori cost functions (f1–f4 in Sec. 3.3/App. C.3) with constraints (C_prec, C_kin, C_col) that do not encode the evaluated success metrics. The RL policy is trained with a fixed dense reward (negative L2 distance to goal, Sec. 4.3), domain randomization with 3 mm noise, and residual actions; it is not fitted to the reported 80% step success or to the multi-step completion rates. The OOD generalist evaluation in Sec. 5.4 holds out the test assembly from training, so those results are genuine transfers rather than construction. The ASAP baseline is from the same group, but it is used as a comparative baseline, not as justification for Fabrica's own capabilities. The only reused prior work is Assemble-Them-All [10] for disassembly-path feasibility inside precedence planning; this is a component-level tool with stated assumptions, and the central claims do not reduce to it. The gap between the abstract's 'complete assembly without human demonstrations' and Table 5's near-zero zero-intervention completion rates is a claim-versus-evidence consistency concern, not a circularity in the derivation. Accordingly, no circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the explicitly stated A1-A5 scope restrictions, the reliability of the authors' prior physics-based disassembly planner, the sufficiency of 3 mm domain randomization for sim-to-real transfer, and the accuracy of the simulators. No new physical entities are postulated. The evaluation protocol (three trials per step, up to two human interventions) is a free parameter that inflates the headline step success and enables the 'complete assembly' claim.

free parameters (3)
  • Domain randomization noise amplitude = 3 mm
    Chosen for training; the paper acknowledges real-world errors are much larger than 3 mm, and this directly affects the policy transfer claim.
  • Number of trials per step = 3
    The headline 80% step success in Table 4 appears to include up to three attempts per step; single-trial numbers in Appendix E.2 are lower, so the central success metric depends on this protocol.
  • PLAI action scale and error threshold = 0.001 and 0.02
    Manually tuned deployment parameters, held constant across assemblies; they affect action consistency and success rates.
assumptions (4)
  • ad hoc to paper Scope restrictions A1-A5: insertion-only assembly, no subassembly reorientation, monotonic assembly (no regrasps or handovers), no force/torque constraints, finite grasp set per part.
    These assumptions define the problem boundary and are explicitly stated in Sec 7; they exclude screwing, sliding, heavy parts, and reorientation, which limits the generality of the assembly capability.
  • domain assumption The physics-based disassembly planner from Assemble-Them-All [10] finds all feasible disassembly paths within its timeout.
    The precedence tiers and graph in Sec 3.1 rely on this planner to decide which parts can be removed; if it misses feasible directions, the sequence search and grasp filtering could produce invalid or suboptimal plans.
  • ad hoc to paper Straight-line insertion paths are sufficient for all assembly steps in the benchmark, and the path-centric transformation in Sec 4.1 is valid for these paths.
    The policy's equivariant coordinate frame maps straight-line assembly motions to top-down insertions; curved or complex mating paths would not be covered by this transformation.
  • domain assumption RedMax and Isaac Gym simulate contact dynamics accurately enough for zero-shot sim-to-real transfer under 3 mm randomization and residual actions.
    The paper's own failure analysis (Sec 5.4) notes large sim-to-real gaps, including unmodeled holding gripper and larger-than-simulated errors, so this assumption is only partially satisfied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning." pith.science (2026). https://pith.science/paper/M4MUSVHP

@misc{pith2026250605168,
  author       = {Pith},
  title        = {Pith review of: Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4MUSVHP}},
  note         = {Machine review of arXiv:2506.05168}
}
read the original abstract

Multi-part assembly poses significant challenges for robots to execute long-horizon, contact-rich manipulation with generalization across complex geometries. We present Fabrica, a dual-arm robotic system capable of end-to-end planning and control for autonomous assembly of general multi-part objects. For planning over long horizons, we develop hierarchies of precedence, sequence, grasp, and motion planning with automated fixture generation, enabling general multi-step assembly on any dual-arm robots. The planner is made efficient through a parallelizable design and is optimized for downstream control stability. For contact-rich assembly steps, we propose a lightweight reinforcement learning framework that trains generalist policies across object geometries, assembly directions, and grasp poses, guided by equivariance and residual actions obtained from the plan. These policies transfer zero-shot to the real world and achieve 80% successful steps. For systematic evaluation, we propose a benchmark suite of multi-part assemblies resembling industrial and daily objects across diverse categories and geometries. By integrating efficient global planning and robust local control, we showcase the first system to achieve complete and generalizable real-world multi-part assembly without domain knowledge or human demonstrations. Project website: http://fabrica.csail.mit.edu/

Figures

Figures reproduced from arXiv: 2506.05168 by the authors.

Figure 1
Figure 1. Our proposed dual-arm robotic system demonstrates adaptive manipulation and assembly [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. System overview. Fabrica takes part meshes and hardware configurations as inputs. It [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Top: benchmark assemblies. Bottom: the auto-generated pickup fixtures in Sec. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 1
Figure 1. Figure 1: Step-by-step rendered assembly executions on different assemblies with different robots. [PITH_FULL_IMAGE:figures/full_fig_p014_1.png]
Figure 2
Figure 2. Figure 2: Step-by-step real-world assembly executions on different assemblies with Panda robots, [PITH_FULL_IMAGE:figures/full_fig_p015_2.png]
Figure 3
Figure 3. Figure 3: Physical setup for integerat￾ing vision feedback. Left: Camera de￾tails. Right: The mounted configuration on the robot wrist. The vision integration requires only RGB input without high imaging quality, allowing for the use of low-cost cameras. We utilize an Arducam B0…
Figure 4
Figure 4. Figure 4: Example outputs from VLM during corrective alignment. The VLM identifies spatial [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 43 canonical work pages

  1. [1]

    J. Xu, S. Kim, T. Chen, A. R. Garcia, P. Agrawal, W. Matusik, and S. Sueda. Efficient tactile simulation with differentiability for robotic manipulation. In6th Annual Conference on Robot Learning, 2022

    work page 2022

  2. [2]

    Zhang, M

    X. Zhang, M. Tomizuka, and H. Li. Bridging the sim-to-real gap with dynamic compliance tuning for industrial insertion. InICRA, 2024

    work page 2024

  3. [3]

    B. Tang, I. Akinola, J. Xu, B. Wen, A. Handa, K. Van Wyk, D. Fox, G. S. Sukhatme, F. Ramos, and Y . Narang. Automate: Specialist and generalist assembly policies over diverse geometries. InRobotics: Science and Systems, 2024

    work page 2024

  4. [4]

    Halperin, J.-C

    D. Halperin, J.-C. Latombe, and R. H. Wilson. A general framework for assembly planning: The motion space approach.Algorithmica, 26(3):577–601, 2000

    work page 2000

  5. [5]

    Sundaram, I

    S. Sundaram, I. Remmler, and N. M. Amato. Disassembly sequencing using a motion plan- ning approach. InProceedings 2001 ICRA. IEEE International Conference on Robotics and Automation (Cat. No. 01CH37164), volume 2, pages 1475–1480. IEEE, 2001

    work page 2001

  6. [6]

    D. T. Le, J. Cort ´es, and T. Sim´eon. A path planning approach to (dis) assembly sequencing. In 2009 IEEE International Conference on Automation Science and Engineering, pages 286–291. IEEE, 2009

    work page 2009

  7. [7]

    Zhang, R

    X. Zhang, R. Belfer, P. G. Kry, and E. V ouga. C-space tunnel discovery for puzzle path plan- ning.ACM Transactions on Graphics (TOG), 39(4):104–1, 2020

    work page 2020

  8. [8]

    N. Funk, G. Chalvatzaki, B. Belousov, and J. Peters. Learn2assemble with structured repre- sentations and search for robotic architectural construction. InConference on Robot Learning, pages 1401–1411. PMLR, 2022

    work page 2022

Show all 56 references
  1. [9]

    S. K. S. Ghasemipour, S. Kataoka, B. David, D. Freeman, S. S. Gu, and I. Mordatch. Blocks assemble! learning to assemble with large-scale structured reinforcement learning. InInterna- tional Conference on Machine Learning, pages 7435–7469. PMLR, 2022

    2022

  2. [10]

    Y . Tian, J. Xu, Y . Li, J. Luo, S. Sueda, H. Li, K. D. Willis, and W. Matusik. Assemble them all: Physics-based planning for generalizable assembly by disassembly.ACM Transactions on Graphics (TOG), 41(6):1–11, 2022

    2022

  3. [11]

    Y . Tian, K. D. Willis, B. Al Omari, J. Luo, P. Ma, Y . Li, F. Javid, E. Gu, J. Jacob, S. Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380...

    2024

  4. [12]

    Rodrıguez, K

    I. Rodrıguez, K. Nottensteiner, D. Leidner, M. Kaßecker, F. Stulp, and A. Albu-Sch ¨affer. Iter- atively refined feasibility checks in robotic assembly sequence planning.IEEE Robotics and Automation Letters, 4(2):1416–1423, 2019

    2019

  5. [13]

    X. Zhu, D. K. Jha, D. Romeres, L. Sun, M. Tomizuka, and A. Cherian. Multi-level reasoning for robotic assembly: From sequence inference to contact selection. In2024 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 816–823. IEEE, 2024. 10

    2024

  6. [14]

    Su ´arez-Ruiz, X

    F. Su ´arez-Ruiz, X. Zhou, and Q.-C. Pham. Can robots assemble an ikea chair?Science Robotics, 3(17):eaat6385, 2018

    2018

  7. [15]

    N ¨agele, A

    L. N ¨agele, A. Hoffmann, A. Schierl, and W. Reif. Legobot: Automated planning for coordi- nated multi-robot assembly of lego structures. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9088–9095. IEEE, 2020

    2020

  8. [16]

    Huang, C

    Y . Huang, C. R. Garrett, I. Ting, S. Parascho, and C. T. Mueller. Robotic additive construction of bar structures: Unified sequence and motion planning.Construction Robotics, 5:115–130, 2021

    2021

  9. [17]

    Huang, P

    Y . Huang, P. Y . V . Leung, C. Garrett, F. Gramazio, M. Kohler, and C. Mueller. The new analog: A protocol for linking design and construction intent with algorithmic planning for robotic assembly of complex structures. InProceedings of the 6th Annual ACM Symposium on Computa...

    2021

  10. [18]

    Z. Wang, F. Kennel-Maushart, Y . Huang, B. Thomaszewski, and S. Coros. A temporal coherent topology optimization approach for assembly planning of bespoke frame structures.ACM Transactions on Graphics (TOG), 42(4):1–13, 2023

    2023

  11. [19]

    Liang, S.-C

    C.-J. Liang, S.-C. Kang, and M.-H. Lee. Ras: a robotic assembly system for steel structure erection and assembly.International Journal of Intelligent Robotics and Applications, 1:459– 476, 2017

    2017

  12. [20]

    D ¨orfler, T

    K. D ¨orfler, T. Sandy, M. Giftthaler, F. Gramazio, M. Kohler, and J. Buchli. Mobile robotic brickwork: automation of a discrete robotic fabrication process using an autonomous mobile robot.Robotic Fabrication in Architecture, Art and Design 2016, pages 204–217, 2016

    2016

  13. [21]

    A. A. Apolinarska, M. Pacher, H. Li, N. Cote, R. Pastrana, F. Gramazio, and M. Kohler. Robotic assembly of timber joints using reinforcement learning.Automation in Construction, 125:103569, 2021

    2021

  14. [22]

    Kramberger, A

    A. Kramberger, A. Kunic, I. Iturrate, C. Sloth, R. Naboni, and C. Schlette. Robotic assembly of timber structures in a human-robot collaboration setup.Frontiers in Robotics and AI, 8: 768038, 2022

    2022

  15. [23]

    Thomas, M

    G. Thomas, M. Chien, A. Tamar, J. A. Ojea, and P. Abbeel. Learning robotic assembly from cad. In2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3524–

  16. [24]

    Y . Fan, J. Luo, and M. Tomizuka. A learning framework for high precision industrial assembly. In2019 International Conference on Robotics and Automation (ICRA), pages 811–817. IEEE, 2019

    2019

  17. [25]

    Narang, K

    Y . Narang, K. Storey, I. Akinola, M. Macklin, P. Reist, L. Wawrzyniak, Y . Guo, A. Mora- vanszky, G. State, M. Lu, et al. Factory: Fast contact for robotic assembly.arXiv preprint arXiv:2205.03532, 2022

    2022 arXiv

  18. [26]

    B. Tang, M. A. Lin, I. Akinola, A. Handa, G. S. Sukhatme, F. Ramos, D. Fox, and Y . Narang. Industreal: Transferring contact-rich assembly tasks from simulation to reality.arXiv preprint arXiv:2305.17110, 2023

    2023 arXiv

  19. [27]

    Zhang, C

    X. Zhang, C. Wang, L. Sun, Z. Wu, X. Zhu, and M. Tomizuka. Efficient sim-to-real transfer of contact-rich manipulation skills with online admittance residual learning. InConference on Robot Learning, pages 1621–1639. PMLR, 2023

    2023

  20. [28]

    J. Luo, Z. Hu, C. Xu, Y . L. Tan, J. Berg, A. Sharma, S. Schaal, C. Finn, A. Gupta, and S. Levine. Serl: A software suite for sample-efficient robotic reinforcement learning.arXiv preprint arXiv:2401.16013, 2024. 11

    2024 arXiv

  21. [29]

    Noseworthy, B

    M. Noseworthy, B. Tang, B. Wen, A. Handa, N. Roy, D. Fox, F. Ramos, Y . Narang, and I. Aki- nola. Forge: Force-guided exploration for robust contact-rich manipulation under uncertainty. arXiv preprint arXiv:2408.04587, 2024

    2024 arXiv

  22. [30]

    Ankile, A

    L. Ankile, A. Simeonov, I. Shenfeld, and P. Agrawal. Juicer: Data-efficient imitation learning for robotic assembly.arXiv, 2024

    2024

  23. [31]

    Ankile, A

    L. Ankile, A. Simeonov, I. Shenfeld, M. Torne, and P. Agrawal. From imitation to refinement – residual rl for precise assembly, 2024. URLhttps://arxiv.org/abs/2407.16677

    2024 arXiv

  24. [32]

    Huang, D

    H. Huang, D. Wang, A. Tangri, R. Walters, and R. Platt. Leveraging symmetries in pick and place.The International Journal of Robotics Research, page 02783649231225775, 2024

    2024

  25. [33]

    Simeonov, Y

    A. Simeonov, Y . Du, A. Tagliasacchi, J. B. Tenenbaum, A. Rodriguez, P. Agrawal, and V . Sitz- mann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394–6400. IEEE, 2022

    2022

  26. [34]

    Simeonov, Y

    A. Simeonov, Y . Du, Y .-C. Lin, A. R. Garcia, L. P. Kaelbling, T. Lozano-P´erez, and P. Agrawal. Se (3)-equivariant relational rearrangement with neural descriptor fields. InConference on Robot Learning, pages 835–846. PMLR, 2023

    2023

  27. [35]

    J. Yang, C. Deng, J. Wu, R. Antonova, L. Guibas, and J. Bohg. Equivact: Sim (3)-equivariant visuomotor policies beyond rigid object manipulation.arXiv preprint arXiv:2310.16050, 2023

    2023 arXiv

  28. [36]

    D. Wang, S. Hart, D. Surovik, T. Kelestemur, H. Huang, H. Zhao, M. Yeatman, J. Wang, R. Walters, and R. Platt. Equivariant diffusion policy.arXiv preprint arXiv:2407.01812, 2024

    2024 arXiv

  29. [37]

    J. Seo, N. P. Prakash, X. Zhang, C. Wang, J. Choi, M. Tomizuka, and R. Horowitz. Contact- rich se (3)-equivariant robot manipulation task learning via geometric impedance control.IEEE Robotics and Automation Letters, 2023

    2023

  30. [38]

    W. Lian, T. Kelch, D. Holz, A. Norton, and S. Schaal. Benchmarking off-the-shelf solutions to robotic assembly tasks. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1046–1053. IEEE, 2021

    2021

  31. [39]

    L. Fan, Y . Zhu, J. Zhu, Z. Liu, O. Zeng, A. Gupta, J. Creus-Costa, S. Savarese, and L. Fei-Fei. Surreal: Open-source reinforcement learning framework and robot manipulation benchmark. InConference on robot learning, pages 767–782. PMLR, 2018

    2018

  32. [40]

    Y . Lee, E. S. Hu, and J. J. Lim. Ikea furniture assembly environment for long-horizon complex manipulation tasks. In2021 ieee international conference on robotics and automation (icra), pages 6343–6349. IEEE, 2021

    2021

  33. [41]

    M. Heo, Y . Lee, D. Lee, and J. J. Lim. Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. InRobotics: Science and Systems, 2023

    2023

  34. [42]

    J. Luo, C. Xu, F. Liu, L. Tan, Z. Lin, J. Wu, P. Abbeel, and S. Levine. Fmb: a functional manipulation benchmark for generalizable robotic learning.arXiv preprint arXiv:2401.08553, 2024

    2024 arXiv

  35. [43]

    Alami, J.-P

    R. Alami, J.-P. Laumond, and T. Sim ´eon. Two manipulation planning algorithms. InWAFR Proceedings of the workshop on Algorithmic foundations of robotics, pages 109–125. AK Pe- ters, Ltd. Natick, MA, USA, 1994

    1994

  36. [44]

    Sim ´eon, J.-P

    T. Sim ´eon, J.-P. Laumond, J. Cort´es, and A. Sahbani. Manipulation planning with probabilistic roadmaps.The International Journal of Robotics Research, 23(7-8):729–746, 2004. 12

    2004

  37. [45]

    J. J. Kuffner and S. M. LaValle. Rrt-connect: An efficient approach to single-query path planning. InProceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065), volume 2, pages 995–1001. IEEE, 2000

    2000

  38. [46]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  39. [47]

    Doshi, M

    K. Doshi, M. Bagatella, and S. Coros. Problem space transformations for generalisation in behavioural cloning.arXiv preprint arXiv:2411.04056, 2024

    2024

  40. [48]

    Silver, K

    T. Silver, K. Allen, J. Tenenbaum, and L. Kaelbling. Residual policy learning.arXiv preprint arXiv:1812.06298, 2018

    2018 arXiv

  41. [49]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State. Isaac gym: High performance gpu-based physics simula- tion for robot learning, 2021

    2021

  42. [50]

    Makoviichuk and V

    D. Makoviichuk and V . Makoviychuk. rl-games: A high-performance framework for rein- forcement learning.https://github.com/Denys88/rl_games, May 2021

    2021

  43. [51]

    A. Lodi, S. Martello, and M. Monaci. Two-dimensional packing problems: A survey.European journal of operational research, 141(2):241–252, 2002

    2002

  44. [52]

    Jyl ¨anki

    J. Jyl ¨anki. A thousand ways to pack the bin-a practical approach to two-dimensional rectangle bin packing, 2010. URLhttps://raw.githubusercontent.com/juj/ RectangleBinPack/master/RectangleBinPack.pdf. 13 Appendix A Qualitative Results Fig. 1 and Fig. 2 show qualitative resul...

    2010

  45. [53]

    We use the Maximal Rectangles algorithm [52] to pack the bounding boxes into an initial bounding area

  46. [54]

    We check the collisions between the gripper and the precedent parts of the grasped part

  47. [55]

    If any collisions are detected, the colliding parts are buffered with additional spacing and the bin-packing process is performed again

  48. [56]

    move right

    We increase the bin area once it is not enough to find a packing solution given the increased rectangle sizes. Once an optimal packing configuration is determined, the fixture is generated by creating mold cav- ities that accommodate the part shapes. A minimal mold depth is ca...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.