pith. sign in

arxiv: 2607.01212 · v1 · pith:UICTJJHYnew · submitted 2026-07-01 · 💻 cs.RO · cs.AI

FurnitureVLA: Learning Long-Horizon Bimanual Furniture Assembly with Vision-Language-Action Model

Chenyang Ma , Yue Yang , Radu Corcodel , Siddarth Jain , Andrew Wu , Chiori Hori , Diego Romeres This is my paper

Pith reviewed 2026-07-02 10:47 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords vision-language-action modelsbimanual manipulationlong-horizon tasksfurniture assemblyprogress predictionsim-to-real transferrobot learning
0
0 comments X

The pith

A vision-language-action model that jointly predicts actions and progress reaches 80 percent success on long-horizon bimanual furniture assembly

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

FurnitureVLA is the first systematic study of real-scale bimanual furniture assembly with vision-language-action models. The authors build a scalable simulation pipeline for expert data and evaluation together with a VR teleoperation system that lets one operator collect high-quality bimanual demonstrations. They finetune a progress-enhanced VLA on semantically grounded subtasks so the model outputs both control actions and a continuous progress signal. This combination produces automatic subtask transitions and limits error accumulation across sequences of up to 1550 steps. Average simulation success rises from 48 to 80 percent across three furniture types, with real Kinova Gen3 validation showing only a 16 percent drop on the hardest task.

Core claim

Finetuning a vision-language-action model on semantically grounded subtasks to jointly predict robot actions and a continuous progress signal allows automatic subtask transitions and reduces compounding errors during inference. The resulting system handles extreme long-horizon assembly with up to seven subtasks and 1550 control steps. It achieves 80 percent average simulation success across three furniture types, up from 48 percent for baselines, plus an extra 21 percent gain from perception and control design choices, and transfers to a real dual-arm Kinova Gen3 platform with only a 16 percent performance drop on the most difficult task.

What carries the argument

progress-enhanced VLA finetuned to jointly output actions and a continuous progress signal for each semantically grounded subtask

If this is right

  • Automatic subtask transitions occur without separate segmentation or handoff logic.
  • Compounding errors are reduced across sequences of up to 1550 control steps.
  • Average simulation success reaches 80 percent across three different furniture types.
  • Studied perception and control design factors deliver an additional 21 percent performance gain.
  • Real-world transfer to the Kinova Gen3 platform incurs only a 16 percent drop on the hardest task.

Where Pith is reading between the lines

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

  • The same joint action-plus-progress prediction could support recovery behaviors when small execution errors occur.
  • The VR single-operator collection method may scale to gather data for other long-horizon bimanual tasks.
  • Defining analogous progress signals for different domains could extend the approach beyond furniture assembly.

Load-bearing premise

The simulation pipeline and VR-collected demonstrations produce data whose distribution is close enough to real Kinova Gen3 execution that sim-to-real transfer remains reliable even after 1550 control steps.

What would settle it

Execute the trained model on the physical Kinova Gen3 robot for a complete furniture assembly sequence exceeding 1550 steps while introducing small visual or positional perturbations and check whether success on the hardest task falls below 64 percent.

Figures

Figures reproduced from arXiv: 2607.01212 by Andrew Wu, Chenyang Ma, Chiori Hori, Diego Romeres, Radu Corcodel, Siddarth Jain, Yue Yang.

Figure 1
Figure 1. Figure 1: Real-scale bimanual furniture assembly with Vision-Language-Action models. We introduce FurnitureVLA, the first systematic study of this challenging setting. It comprises a scalable simulation pipeline for data generation and evaluation, and a tailored VR teleoperation system for high-quality demonstration collection. We propose a progress-enhanced VLA to tackle long-horizon assembly. Abstract— Current wor… view at source ↗
Figure 2
Figure 2. Figure 2: Bimanual furniture assembly tasks. These tasks require executing diverse manipulation skills over extremely long horizons. A simple item (LACK side table) requires 12 skill executions (650 steps), while a complex assembly (IVAR chair) requires 25 (1550 steps). Assumptions. We make two key assumptions. First, we use magnets to attach furniture parts once they are aligned, bypassing the screwing process, whi… view at source ↗
Figure 3
Figure 3. Figure 3: System setup. The teleoperator wears a Meta Quest 3 headset at the neck to track hand poses. A green screen reduces visual noise; a white cover occludes the teleoperator. Bottom-right: tabletop schematic with symmetric dual Kinova placement. skills, such as grasping, alignment, insertion, lifting, and rotation. As a result, a simple assembly (e.g., LACK with 4 subtasks) involves roughly 12 skill executions… view at source ↗
Figure 5
Figure 5. Figure 5: Subtask success rates (monotonically decreasing). TABLE I: Assembly performance (success rates ↑). Method LACK KALLAX IVAR Average π0.5 (zero-shot) 0.00 0.00 0.00 0.00 π0.5 (monolithic finetuned) 0.91 0.11 0.41 0.48 FurnitureVLA 0.98 0.85 0.56 0.80 are free of contact and force constraints. Small errors are less likely to amplify, yielding a narrower, more consistent initial￾state distribution and reducing… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative real-robot rollouts. (Left) successful long-horizon assembly of the IVAR chair. (Right) emergent self-correction behaviors: regrasping for improved contact and fine-grained magnet alignment. 10.00 mm 15.13 mm 10° 36.69 mm 13.31 mm [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real-world precision requirements. (Left) assembly tolerance with magnets. (Right) gripper clearance and multi-magnet alignment challenges. VI. REAL-WORLD EXPERIMENTS We validate our approach on a real bimanual Kinova Gen3 platform. We test on the most challenging IVAR chair. A. Implementation Details We use Quest2ROS [46] to map VR controller poses and inputs to robot end-effector commands. Colored marker… view at source ↗
Figure 9
Figure 9. Figure 9: Simulation and real-world observation examples. (Left) simulation rear-view RGB and depth observations. (Right) real￾world observations, including a rear-view image captured by an Intel RealSense D435 and two wrist-mounted images from the Kinova arms. All images are raw and unprocessed. IVAR Chair Two-Stage Simulation. Isaac Gym does not support runtime weld constraints, so once the left side frame, rails,… view at source ↗
Figure 11
Figure 11. Figure 11: Predefined orientation presets. Left to right: grasp poses for rail grasping, rail attachment, seat panel grasping, and dual-arm coordinated grasping of the left side frame. via a simple proportional (P) gain, and execute it through the same Kinova velocity controller. Teleoperation Control Mapping. Since the teleoperator stands opposite the dual-arm system, control is defined in the teleoperator’s frame … view at source ↗
Figure 12
Figure 12. Figure 12: Subtask success rates under different design choices. Results correspond to Table II and Table III [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
read the original abstract

Current work on robot furniture assembly mostly focuses on toy-scale settings or single-arm manipulation. We introduce FurnitureVLA, the first systematic study of real-scale bimanual furniture assembly using Vision-Language-Action models (VLAs). We formalize the task, develop a scalable simulation pipeline for expert data generation and evaluation, and build a VR teleoperation system for single-operator bimanual control to collect high-quality real-world demonstrations. To address extreme long-horizon assembly with up to 7 subtasks and 1550 control steps, we propose a progress-enhanced VLA, finetuned on semantically grounded subtasks, that jointly predicts actions and a continuous progress signal, enabling automatic subtask transitions and reducing compounding errors during inference. We further study perception and control design factors that critically affect precision in real-scale assembly. FurnitureVLA improves average simulation success from 48% to 80% compared to baselines across three furniture types, with an additional 21% gain from our design factor study. We validate on a real Kinova Gen3 platform with only 16% drop on the hardest task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces FurnitureVLA as the first systematic study of real-scale bimanual furniture assembly with VLAs. It formalizes the task (up to 7 subtasks, 1550 steps), builds a scalable sim pipeline and VR teleoperation system for data collection, and proposes a progress-enhanced VLA finetuned on semantically grounded subtasks that jointly predicts actions and a continuous progress signal to enable automatic transitions and reduce compounding errors. It reports average simulation success rising from 48% to 80% across three furniture types (plus 21% from design-factor study) and validates on a real Kinova Gen3 with a 16% drop on the hardest task.

Significance. If the reported gains and sim-to-real transfer hold under rigorous evaluation, the work would be significant as one of the first demonstrations of VLAs scaling to long-horizon, real-scale bimanual assembly with hardware validation. The joint action+progress formulation and design-factor analysis address practical challenges in compounding errors and precision that are load-bearing for this domain.

major comments (2)
  1. [Abstract] Abstract: the central empirical claim (48% o 80% simulation success, 16% real drop) is presented without error bars, trial counts, baseline implementation details, data volumes, or ablation numbers isolating the progress signal's contribution; this prevents assessment of whether the result is robust or sensitive to post-hoc choices.
  2. [Abstract] Abstract (real-robot validation paragraph): the claim that sim-to-real transfer remains reliable after 1550 steps rests on the unquantified assumption that VR demonstrations and the simulation pipeline produce trajectories whose state, visual, and dynamics distributions match real Kinova Gen3 execution; no alignment metrics (state visitation divergence, progress-signal calibration error, or failure-mode breakdown) are supplied to support this.
minor comments (1)
  1. [Abstract] The abstract states 'an additional 21% gain from our design factor study' without indicating whether this is additive to the 80% or measured separately; clarify the exact comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that greater transparency on experimental details would strengthen the presentation and will revise the abstract accordingly while preserving its conciseness. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim (48% o 80% simulation success, 16% real drop) is presented without error bars, trial counts, baseline implementation details, data volumes, or ablation numbers isolating the progress signal's contribution; this prevents assessment of whether the result is robust or sensitive to post-hoc choices.

    Authors: The abstract is intentionally concise, but we acknowledge that key statistics would aid immediate assessment. The full manuscript reports these details in Section 4 (Experiments): error bars and trial counts (50 independent trials per furniture type), data volumes (Section 3.2), baseline implementations (Section 3), and ablations isolating the progress signal (Section 4.3, showing its isolated contribution). We will revise the abstract to include a brief qualifier such as "across 50 trials per task (with standard error)" and a parenthetical note on the ablation results. Full baseline code and hyperparameter details remain in the methods section due to space limits. revision: yes

  2. Referee: [Abstract] Abstract (real-robot validation paragraph): the claim that sim-to-real transfer remains reliable after 1550 steps rests on the unquantified assumption that VR demonstrations and the simulation pipeline produce trajectories whose state, visual, and dynamics distributions match real Kinova Gen3 execution; no alignment metrics (state visitation divergence, progress-signal calibration error, or failure-mode breakdown) are supplied to support this.

    Authors: Sections 3.1–3.2 describe the VR teleoperation and simulation pipeline explicitly engineered to match the Kinova Gen3 kinematics, visuals, and control interface. The primary evidence of transfer is the reported 16% performance drop on the hardest task (Section 4.4), which includes qualitative failure-mode analysis. We do not currently supply quantitative alignment metrics such as state visitation divergence or progress-signal calibration error. We will revise the abstract and/or real-robot subsection to explicitly acknowledge this assumption and add a short discussion of the design choices that support distributional similarity. Adding new quantitative metrics would require additional post-hoc analysis of existing trajectories. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical VLA performance lift is self-contained

full rationale

The paper reports an empirical improvement (48% to 80% simulation success) from a progress-enhanced VLA trained on VR demonstrations and evaluated in simulation plus real hardware. No equations, fitted parameters, or derivation chain are presented that reduce a claimed prediction to its own inputs by construction. The central result is a measured performance delta on held-out tasks rather than a self-referential mathematical claim, and no load-bearing self-citation or uniqueness theorem is invoked in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no equations, parameters, or background assumptions; ledger left empty pending full text.

pith-pipeline@v0.9.1-grok · 5746 in / 1075 out tokens · 23977 ms · 2026-07-02T10:47:25.170886+00:00 · methodology

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

50 extracted references · 8 canonical work pages · 3 internal anchors

  1. [1]

    Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation,

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

    2023

  2. [2]

    Generalize by touching: Tactile ensemble skill transfer for robotic furniture assembly,

    H. Lin, R. Corcodel, and D. Zhao, “Generalize by touching: Tactile ensemble skill transfer for robotic furniture assembly,” inInternational Conference on Robotics and Automation, 2024

    2024

  3. [3]

    JUICER: data- efficient imitation learning for robotic assembly,

    L. Ankile, A. Simeonov, I. Shenfeld, and P. Agrawal, “JUICER: data- efficient imitation learning for robotic assembly,” inInternational Conference on Intelligent Robots and Systems, 2024

    2024

  4. [4]

    Fabrica: Dual-arm assembly of general multi-part objects via integrated planning and learning,

    Y . Tian, J. Jacob, Y . Huang, J. Zhao, E. Gu, P. Ma, A. Zhang, F. Javid, B. Romero, S. Chittaet al., “Fabrica: Dual-arm assembly of general multi-part objects via integrated planning and learning,” inConference on Robot Learning, 2025

    2025

  5. [5]

    Efficient sim-to-real transfer of contact-rich manipulation skills with online admittance residual learning,

    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, 2023

    2023

  6. [6]

    From imitation to refinement - residual rl for precise assembly,

    L. Ankile, A. Simeonov, I. Shenfeld, M. Torne, and P. Agrawal, “From imitation to refinement - residual rl for precise assembly,” in International Conference on Robotics and Automation, 2025

    2025

  7. [7]

    FORGE: force-guided exploration for robust contact-rich manipulation under uncertainty,

    M. Noseworthy, B. Tang, B. Wen, A. Handa, C. C. Kessens, N. Roy, D. Fox, F. Ramos, Y . Narang, and I. Akinola, “FORGE: force-guided exploration for robust contact-rich manipulation under uncertainty,” Robotics and Automation Letters, 2025

    2025

  8. [8]

    Vt-refine: Learning bimanual assembly with visuo-tactile feedback via simulation fine-tuning,

    B. Huang, J. Xu, I. Akinola, W. Yang, B. Sundaralingam, R. O’Flaherty, D. Fox, X. Wang, A. Mousavian, Y . Chao, and Y . Li, “Vt-refine: Learning bimanual assembly with visuo-tactile feedback via simulation fine-tuning,” inConference on Robot Learning, 2025

    2025

  9. [9]

    Mobile ALOHA: learning bimanual mobile manipulation using low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile ALOHA: learning bimanual mobile manipulation using low-cost whole-body teleoperation,” in Conference on Robot Learning, 2024

    2024

  10. [10]

    Robot- mover: Learning to move large objects from human demonstrations,

    T. Li, J. Truong, J. Yang, A. Clegg, A. Rai, S. Ha, and X. Puig, “Robot- mover: Learning to move large objects from human demonstrations,” arXiv preprint arXiv:2502.05271, 2025

    work page arXiv 2025

  11. [11]

    Openvla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “Openvla: An open-source vision-language-action model,” inConference on Robot Learning, 2024

    2024

  12. [12]

    Gen-0: Embodied foundation models that scale with physical interaction,

    Generalist AI Team, “Gen-0: Embodied foundation models that scale with physical interaction,” https://generalistai.com/blog/ nov-04-2025-GEN-0, 2025

    2025

  13. [13]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. J. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in International Conference on Artificial Intelligence and Statistics, 2011

    2011

  14. [14]

    BOSS: benchmark for observation space shift in long-horizon task,

    Y . Yang, L. Zhao, M. Ding, G. Bertasius, and D. Szafir, “BOSS: benchmark for observation space shift in long-horizon task,”Robotics and Automation Letters, 2025

    2025

  15. [15]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inRobotics: Science and Systems, 2023

    2023

  16. [16]

    Can robots assemble an IKEA chair?

    X. Zhou and Q. Pham, “Can robots assemble an IKEA chair?”Science Robotics, 2018

    2018

  17. [17]

    Manual2skill: Learning to read manuals and acquire robotic skills for furniture assembly using vision-language models,

    C. Tie, S. Sun, J. Zhu, Y . Liu, J. Guo, Y . Hu, H. Chen, J. Chen, R. Wu, and L. Shao, “Manual2skill: Learning to read manuals and acquire robotic skills for furniture assembly using vision-language models,” in Robotics: Science and Systems, 2025

    2025

  18. [18]

    Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors,

    C. Ma, K. Lu, T.-Y . Cheng, N. Trigoni, and A. Markham, “Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors,” inNeural Information Processing Systems, 2024

    2024

  19. [19]

    IKEA furniture assembly environ- ment for long-horizon complex manipulation tasks,

    Y . Lee, E. S. Hu, and J. J. Lim, “IKEA furniture assembly environ- ment for long-horizon complex manipulation tasks,” inInternational Conference on Robotics and Automation, 2021

    2021

  20. [20]

    Industreal: Transferring contact-rich assembly tasks from simulation to reality,

    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,” inRobotics: Science and Systems, 2023

    2023

  21. [21]

    DECAF: a discrete-event based collaborative human-robot framework for furniture assembly,

    G. Giacomuzzo, M. Terreran, S. Jain, and D. Romeres, “DECAF: a discrete-event based collaborative human-robot framework for furniture assembly,” inInternational Conference on Intelligent Robots and Systems, 2024

    2024

  22. [22]

    Arch: Hierarchical hybrid learning for long-horizon contact-rich robotic assembly,

    J. Sun, A. Curtis, Y . You, Y . Xu, M. Koehle, Q. Chen, S. Huang, L. Guibas, S. Chitta, M. Schwageret al., “Arch: Hierarchical hybrid learning for long-horizon contact-rich robotic assembly,” inConference on Robot Learning, 2025

    2025

  23. [23]

    Coopera: Continual open-ended human-robot assistance,

    C. Ma, K. Lu, R. Desai, X. Puig, A. Markham, and N. Trigoni, “Coopera: Continual open-ended human-robot assistance,” inAdvances in Neural Information Processing Systems, 2025

    2025

  24. [24]

    Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation,

    Y . Fan, P. Ding, S. Bai, X. Tong, Y . Zhu, H. Lu, F. Dai, W. Zhao, Y . Liu, S. Huang, Z. Fan, B. Chen, and D. Wang, “Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation,” inConference on Robot Learning, 2025

    2025

  25. [25]

    Seqvla: Sequential task execution for long-horizon manipulation with completion-aware vision- language-action model,

    R. Yang, Z. An, L. Zhou, and Y . Feng, “Seqvla: Sequential task execution for long-horizon manipulation with completion-aware vision- language-action model,”arXiv preprint arXiv:2509.14138, 2025

    work page arXiv 2025

  26. [26]

    LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies, 2026

    Y . Yang, S. Cheng, Y . Fang, H. Bharadhwaj, M. Ding, G. Bertasius, and D. Szafir, “Lilo-vla: Compositional long-horizon manipulation via linked object-centric policies,”arXiv preprint arXiv:2602.21531, 2026

    work page arXiv 2026

  27. [27]

    KitchenVLA: Iterative vision-language corrections for robotic execution of human tasks,

    K. Lu, C. Ma, C. Hori, and D. Romeres, “KitchenVLA: Iterative vision-language corrections for robotic execution of human tasks,” inInternational Conference on Robotics and Automation Workshop Safe-VLM, 2025

    2025

  28. [28]

    Cyclevla: Proactive self-correcting vision-language-action models via subtask backtracking and minimum bayes risk decoding,

    C. Ma, G. Yang, K. Lu, S. Xu, B. Byrne, N. Trigoni, and A. Markham, “Cyclevla: Proactive self-correcting vision-language-action models via subtask backtracking and minimum bayes risk decoding,”arXiv preprint arXiv:2601.02295, 2026

    work page arXiv 2026

  29. [29]

    Robotic control via embodied chain-of-thought reasoning,

    M. Zawalski, W. Chen, K. Pertsch, O. Mees, C. Finn, and S. Levine, “Robotic control via embodied chain-of-thought reasoning,” inConfer- ence on Robot Learning, 2024

    2024

  30. [30]

    $\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization

    Physical Intelligence, “π0.5: a vision-language-action model with open- world generalization,”arXiv preprint arXiv:2504.16054, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Hi robot: Open-ended instruction following with hierarchical vision-language-action models,

    L. X. Shi, B. Ichter, M. R. Equi, L. Ke, K. Pertsch, Q. Vuong, J. Tanner, A. Walling, H. Wang, N. Fusai, A. Li-Bell, D. Driess, L. Groom, S. Levine, and C. Finn, “Hi robot: Open-ended instruction following with hierarchical vision-language-action models,” inInternational Conference on Machine Learning, 2025

    2025

  32. [32]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    GR00T Team, “GR00T N1: an open foundation model for generalist humanoid robots,”arXiv preprint arXiv:2503.14734, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Open x-embodiment: Robotic learning datasets and RT-X models : Open x-embodiment collaboration,

    Open X-Embodiment Collaboration, “Open x-embodiment: Robotic learning datasets and RT-X models : Open x-embodiment collaboration,” inInternational Conference on Robotics and Automation, 2024

    2024

  34. [34]

    FAST: Efficient Action Tokenization for Vision-Language-Action Models

    Physical Intelligence, “Fast: Efficient action tokenization for vision- language-action models,”arXiv preprint arXiv:2501.09747, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  35. [35]

    Roboarena: Distributed real-world evaluation of generalist robot policies,

    P. Atreya, K. Pertsch, T. Lee, M. J. Kim, A. Jain, A. Kuramshin, C. Eppner, C. Neary, E. Hu, F. Ramoset al., “Roboarena: Distributed real-world evaluation of generalist robot policies,” inConference on Robot Learning, 2025

    2025

  36. [36]

    Openpi comet: Competition solution for 2025 behavior challenge,

    J. Bai, Y .-W. Chao, Q. Chen, J. Gu, M. J. Kim, Z. Li, X. Li, T.-Y . Lin, M.-Y . Liu, N. Maet al., “Openpi comet: Competition solution for 2025 behavior challenge,”arXiv preprint arXiv:2512.10071, 2025

    work page arXiv 2025

  37. [37]

    Screwdriving gripper that mimics human two-handed assembly tasks,

    S. Han, M. Choi, Y . Shin, G. Jang, D. Lee, J. Cho, J. Park, and J. Bae, “Screwdriving gripper that mimics human two-handed assembly tasks,” Robotics, 2022

    2022

  38. [38]

    Robotic fastening with a manual screwdriver,

    L. Tang and Y . Jia, “Robotic fastening with a manual screwdriver,” in International Conference on Robotics and Automation, 2023

    2023

  39. [39]

    CALVIN: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard, “CALVIN: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,”Robotics and Automation Letters, 2022

    2022

  40. [40]

    LIBERO: benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “LIBERO: benchmarking knowledge transfer for lifelong robot learning,” inAdvances in Neural Information Processing Systems, 2023

    2023

  41. [41]

    Robocasa: Large-scale simulation of everyday tasks for generalist robots,

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y . Zhu, “Robocasa: Large-scale simulation of everyday tasks for generalist robots,” inRobotics: Science and Systems, 2024

    2024

  42. [42]

    Isaac gym: High performance GPU based physics simulation for robot learning,

    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 simulation for robot learning,” inNeurIPS Datasets and Benchmarks, 2021

    2021

  43. [43]

    3d warehouse,

    3D Warehouse, “3d warehouse,” https://3dwarehouse.sketchup.com/, 2026

    2026

  44. [44]

    ambientcg,

    ambientCG, “ambientcg,” https://ambientcg.com/, 2026

    2026

  45. [45]

    Blender,

    Blender Foundation, “Blender,” https://www.blender.org, 2025

    2025

  46. [46]

    Quest2ros: An app to facilitate teleoperating robots,

    M. C. Welle, N. Ingelhag, M. Lippi, M. Wozniak, A. Gasparri, and D. Kragic, “Quest2ros: An app to facilitate teleoperating robots,” in Workshop on VAM-HRI, 2024

    2024

  47. [47]

    DROID: A large-scale in-the-wild robot manipulation dataset,

    DROID Team, “DROID: A large-scale in-the-wild robot manipulation dataset,” inRobotics: Science and Systems, 2024. Appendix forFurnitureVLA LACK Side Table KALLAX Shelf IVAR Chair 1234 5 1234 5 1234 5 6 7

    2024

  48. [48]

    Front right leg: 3.0 × 24.0 × 3.0 2) Rear right leg: 3.0 × 24.0 × 3.0 3) Front left leg: 3.0 × 24.0 × 3.0 4) Rear left leg: 3.0 × 24.0 × 3.0 5) Tabletop (base part): 33.0 × 33.0 × 3.0

  49. [49]

    Right side panel: 2.2 × 23.4 × 20.2 2) Center divider: 2.2 × 23.4 × 20.2 3) Left side panel: 2.2 × 23.4 × 20.2 4) Top panel: 45.9 × 2.2 × 23.4 5) Bottom panel (base part): 45.9 × 23.4 × 2.2

  50. [50]

    8:Furniture part taxonomy and geometric properties

    Rear seat rail: 1.2 × 30.0 × 5.8 2) Front seat rail: 1.2 × 30.0 × 5.8 3) Lower back rail: 3.6 × 30.0 × 5.5 4) Upper back rail: 3.6 × 30.0 × 5.5 5) Left side frame: 44.6 × 95.0 × 4.2 6) Seat panel: 36.2 × 40.6 × 1.2 7) Right side frame (stage 1 base part): 44.6 × 95.0 × 4.2 1 234 5 13 24 5 67 435 12 WDHPart Dimensions (W ×D ×H, cm) Fig. 8:Furniture part ta...