Pith. sign in

REVIEW 3 major objections 4 minor 19 references

MEVION is a low-cost, open-source four-arm tabletop data-collection robot that reaches 60 Nm of joint torque and 20.4 rad/s of no-load joint speed, letting researchers collect heavy-object and fast bimanual manipulation data that ALOHA-clas

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 16:28 UTC pith:L3SJJUJB

load-bearing objection An honest, open-source dual-arm hardware platform with strong design and demo evidence; headline torque/speed/payload numbers are motor specs, not measured system limits. the 3 major comments →

arxiv 2607.17970 v1 pith:L3SJJUJB submitted 2026-07-20 cs.RO

MEVION: Low-Cost Open-Source Data Collection System for Powerful and High-Speed Dual-Arm Manipulation

Kento Kawaharazuka , Yoshiki Obinata , Hirokazu Ishida , Jihoon Oh , Temma Suzuki , Shintaro Inoue , Keita Yoneda , Ayumu Iwata
show 1 more author
Kei Okada
This is my paper
classification cs.RO
keywords dual-arm robotdata collectionimitation learningopen-source hardwaresheet metal weldingclosed-link elbowteleoperationlow-cost robotics
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.

This paper introduces MEVION, a tabletop dual-arm data-collection system built to escape the force and speed ceiling of ALOHA, the de facto low-cost standard. Using e-commerce-sourced parts and sheet metal welding that fuses large structural shapes into a handful of components, each 7.0 kg arm reaches a peak joint torque of 60 Nm and a no-load joint speed of 20.4 rad/s, with a full four-arm system costing about USD 14,000. The authors demonstrate teleoperated manipulation of a 3.6 kg dumbbell, a fast frying-pan task, and a traditional Japanese mallet game, then show that imitation learning from 15 demonstrations per task yields over five consecutive autonomous successes. The design moves the elbow motor to the shoulder and drives the lower arm through a closed parallel link, cutting distal mass; software-based gravity compensation replaces mechanical counterweights. The claim is that MEVION makes heavy-object and high-speed manipulation data collection possible at a price and openness that small laboratories can adopt.

Core claim

On its own terms, the paper claims that 'greater force and speed' do not require abandoning the low-cost, open-source formula. MEVION is a four-arm tabletop system with six degrees of freedom per arm, driven by servo motors whose datasheet peak torque is 60 Nm at the shoulder and elbow pitch joints and 17 Nm at the distal joints. The elbow is actuated by a proximal motor through a closed-link parallel mechanism, the same arrangement used in quadruped robots, which removes the elbow motor from the distal arm and lowers the gravity-compensation burden. The metallic structure is reduced to 17 essential parts, 21 including limiters and covers, by welding large sheet-metal shapes into single piec

What carries the argument

Two mechanisms carry the argument. First, the elbow parallel-link drive: a shoulder-mounted motor drives the lower arm through a closed link, as in quadruped robots, so the heavy elbow motor stays proximal and distal mass drops. That is what lets a 60 Nm motor deliver high end-effector torque without a bulky forearm. Second, sheet metal welding: large, complex parts such as the Shoulder-Link, Lower2-Link, and Hand-Slider-Base are fabricated as single welded components, cutting the basic part count from about 31 to 17 and keeping the whole system near USD 14,000. On the software side, gravity compensation is computed by a rigid-body dynamics model and added to a 200 Hz PD controller, replacin

Load-bearing premise

The load-bearing premise is that the published peak torque (60 Nm), no-load speed (20.4 rad/s), and 7.6 kg payload describe what the assembled MEVION arm actually delivers at the end-effector; the paper reports motor datasheet ratings and computed payload rather than a direct force-velocity or static-payload measurement, so if the welded links and closed-link elbow absorb much of that output, the advantage over ALOHA narrows.

What would settle it

Measure MEVION and ALOHA under identical conditions: a static pull test at maximum extension recording end-effector force, a fast no-load swing recording reachable end-effector speed, and a payload test until sag or failure. If MEVION's end-effector force or speed is not above ALOHA's, or if the 60 Nm and 20.4 rad/s figures are only idle motor ratings, the paper's central 'greater force and speed' claim fails.

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

If this is right

  • At USD 3,500 per arm, a research group can collect bimanual manipulation data for tasks requiring sustained force or fast motion, not just slow handling of lightweight objects.
  • The closed-link elbow and software gravity compensation yield a claimed end-effector payload of 7.6 kg at full extension, about 1.7 times OpenArm's 4.5 kg, while keeping the distal arm light.
  • Imitation learning with a transformer-based action-chunking policy works on MEVION with 15 teleoperated demonstrations per task, producing more than five consecutive successes under slight displacements.
  • Because all mechanical parts are orderable online and full design files are released, MEVION is reproducible and modifiable by other laboratories, which could accelerate adoption as an ALOHA alternative.

Where Pith is reading between the lines

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

  • Implicit in the paper's trajectory is a shift in imitation-learning data strategy: rather than fielding many weak robots collecting slow light-object data, a fleet of stronger robots can sample forceful and dynamic manipulations, which are the under-represented tail of current bimanual datasets.
  • A direct head-to-head dataset comparison, collecting the same task on ALOHA and MEVION and training the same policy, would test whether the hardware's extra torque and speed measurably improve downstream policy success on tasks with variable object weights.
  • The closed-link elbow plus sheet-metal-welding recipe could transfer to other morphologies, such as a mobile base or a single-arm variant, with minimal re-engineering.
  • Given the authors' own note that weld quality varies by vendor and is difficult to modify, a simple static-load test of each welded part could be added to the build procedure to keep reproduction consistent across laboratories.

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

3 major / 4 minor

Summary. The paper presents MEVION, a low-cost, open-source, tabletop dual-arm data collection system for imitation learning. The system consists of four 6-DoF arms with parallel-jaw grippers, all-metal sheet-metal-welded structures, proximal elbow motors driving a closed-link parallel mechanism, and RobStride actuators. A unified Python software stack supports MuJoCo simulation, ROS/RViz visualization, and real-world control with software gravity compensation (Eq. (1)). The authors report five teleoperation demonstrations and two ACT-based imitation learning tasks. Headline claims include 60 Nm maximum torque, 20.4 rad/s maximum speed, 7.6 kg payload, and roughly USD 14,000 for the full four-arm system.

Significance. If the performance claims are validated at the system level, MEVION is a useful open-source contribution: it provides a detailed parts list, STEP files, assembly files, software, and a low part count via sheet-metal welding, all orderable through e-commerce. The comparative table with existing open-source systems is a valuable resource. The main shortfall is that the central differentiator — greater force and speed than ALOHA, enabling tasks 'not previously possible' — currently rests on actuator datasheet ratings and computed payload numbers rather than measurements on the assembled robot. Because this differentiator is load-bearing, the manuscript needs additional system-level characterization.

major comments (3)
  1. [§II-B, Table I; §III-A, Figs. 5–6] The maximum torque (60 Nm), maximum speed (20.4 rad/s), and payload (7.6 kg) are presented as MEVION's performance, but Table I's footnote states that the torque and speed are 'the peak torque and no-load joint speed of the highest-output motor' — i.e., actuator datasheet values, not measurements on the assembled arm. The measured joint velocities in Fig. 6 peak at 6.5 rad/s, about one-third of the 20.4 rad/s claim, and the torques in Fig. 5 are described as having 'sufficient margin' from the limits. The assembled arm is affected by structural compliance, bolted joints, the closed-link elbow transmission, and software torque/velocity limits. The paper should provide direct end-effector force/torque measurements or static load tests, joint-speed measurements on the assembled arm, and, if the claim is comparative, corresponding measurements for ALOHA. Without this, the 'greater force and
  2. [§II-B] The statement 'The peak payload at maximum extension is 7.6 kg for MEVION and 4.5 kg for OpenArm' is given without derivation, assumptions, or experimental validation. Since payload capability is central to the heavy-object manipulation claim, the paper should specify the static model used, the worst-case pose, safety factors, and ideally verify with a physical static payload test on the assembled MEVION arm. This is especially important because the demonstrated dumbbell manipulation weighs 3.6 kg (Section III-A), well below the claimed 7.6 kg.
  3. [§III-B; §I] The imitation learning experiments report only that 'the robot successfully executed autonomous actions more than five times consecutively' for two tasks. No success rate, number of trials, or robustness quantification is provided. More importantly, the claim that MEVION enables 'object manipulation tasks not previously possible' is not supported by a comparison with ALOHA or any other existing system on the same tasks. A direct comparison, even a qualitative one showing task failure on ALOHA, would substantiate the central enabling claim. Without it, the statement remains an assertion.
minor comments (4)
  1. [Abstract and Table I] The abstract and Table I state 'maximum torque of 60 Nm' and 'maximum speed 20.4 rad/s' without explicitly noting that these are actuator-level no-load/peak values. Please add a caveat or rephrase to avoid the impression that these are measured assembled-robot end-effector values.
  2. [Fig. 6] The figure labels include 'Maximum No-load Speed: 20.4 rad/s' and 'Maximum Rated Load Speed: 18.9 rad/s' as horizontal lines. Since the measured trace peaks at 6.5 rad/s, the reader may confuse the lines with actual performance. Clarify in the caption that the lines are motor datasheet limits, not achieved speeds.
  3. [Supplementary C, §II-A] The supplementary materials state that 'hand refers to a gripper driven by CyberGear, while hand2 refers to a gripper driven by RobStride01,' but the main text says the hand uses RobStride01. This inconsistency should be resolved.
  4. [Supplementary D] The limitations on sheet-metal welding quality and rework are honestly stated, but the potential impact on joint stiffness and repeatability could be discussed more explicitly in relation to the force/speed claims.

Circularity Check

0 steps flagged

No circularity; the paper's central hardware claim rests on component specifications and physical design, not on a constructed equivalence or fitted prediction.

full rationale

The paper is an engineering demonstration of a new open-source dual-arm data-collection platform. Its central claim of 'greater force and speed' is supported by Table I, whose footnote explicitly identifies the torque and speed values as 'peak torque and no-load joint speed of the highest-output motor' rather than measured end-effector outputs. This is an empirical validation gap: the system-level advantage over ALOHA is not directly demonstrated by end-effector force/speed calibration, and the teleoperation traces in Figs. 5-6 run below the rated limits (maximum measured joint velocity 6.5 rad/s vs. 20.4 rad/s no-load speed). However, this is not circularity. No quantity in the paper is defined in terms of the conclusion; no parameter is fitted and then renamed as a prediction; and the physical mechanisms (closed-link elbow, sheet-metal-welded links, software gravity compensation) are design choices, not results derived from the performance claim. The MEVIUS/MEVIUS2 citations are prior hardware implementations by overlapping authors, but they are cited as examples of an actuation/fabrication approach, not as an external theorem that forces MEVION's specifications. The imitation-learning experiments use independently collected demonstrations and are evaluated by repeated autonomous successes. The lack of direct end-effector force/speed measurement would be a correctness or validation concern, not a circular derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

No free parameters fitted to data. The listed domain assumptions underpin the high-performance claims; no new theoretical entities are introduced.

axioms (2)
  • domain assumption The actuator datasheet values (60 Nm, 20.4 rad/s) translate to the assembled system's output without significant degradation from the parallel-link elbow or welded structure.
    Table I and Figs. 5–6 present these as the system's maximum torque/speed; no end-effector force/speed calibration is shown.
  • domain assumption Software-based gravity compensation using Pinocchio is sufficiently accurate and safe for direct human teleoperation.
    Section II-C and the Limitations section note safety concerns; the authors themselves acknowledge greater safety risk than ALOHA's hardware compensation.

pith-pipeline@v1.3.0-alltime-deepseek · 7314 in / 9633 out tokens · 78591 ms · 2026-08-01T16:28:32.035012+00:00 · methodology

0 comments
read the original abstract

The global competition for developing robotic foundation models is intensifying. Among the data collection systems used for dual-arm robots, ALOHA is representative of being low-cost and open-source, and is widely adopted by researchers as a de facto standard. However, due to its limited ability to generate high forces and speeds, it is difficult to handle heavy objects or perform fast manipulations. To address this, we developed MEVION, a low-cost and open-source dual-arm robot data collection system capable of generating greater force and speed. All parts of this robot can be sourced through e-commerce, and by extensively utilizing sheet metal welding, its large body structure is constructed with a small number of components at low cost, while also simplifying assembly. MEVION is equipped with four 6-DoF arms with parallel grippers. Each arm weighs 7.0 kg and has a maximum torque of 60 Nm, and the entire system can be constructed for about USD 14,000. The elbow joint adopts a closed-link mechanism similar to those used in quadruped robots, which reduces the distal mass and enables higher force and speed output at the end-effector. We demonstrate that MEVION enables data collection for object manipulation tasks not previously possible and supports imitation learning-based motion generation. All hardware and software of this work are included in the Supplementary Materials or https://github.com/haraduka/mevion.

Figures

Figures reproduced from arXiv: 2607.17970 by Ayumu Iwata, Hirokazu Ishida, Jihoon Oh, Kei Okada, Keita Yoneda, Kento Kawaharazuka, Shintaro Inoue, Temma Suzuki, Yoshiki Obinata.

Figure 1
Figure 1. Figure 1: MEVION: Low-cost open-source data collection system for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Design overview of MEVION: The design is divided into the arm and hand, comprising 17 essential metal components in total — 11 in the arm [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Details of sheet metal welding for the Shoulder-Link, Lower2-Link, [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The transition of joint torques during the dumbbell manipulation. [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The transition of joint velocities during the Daruma Otoshi. [PITH_FULL_IMAGE:figures/full_fig_p003_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Teleoperation experiments using MEVION: (a) bottle cap opening, (b) object packing, (c) frying pan operation, (d) 3.6 kg dumbbell manipulation, [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The results of imitation learning using MEVION: (a) towel manipulation and (b) dumbbell packing. [PITH_FULL_IMAGE:figures/full_fig_p004_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: All metal components used in MEVION. C. Software Here, we provide information on the software. All pro￾grams are written in Python and support both MuJoCo-based simulation and real-world control. In addition, visualization in RViz via ROS and interactive visualization using Scikit￾Robot are also supported. Accordingly, the package includes XML model files for MuJoCo, URDF model files for RViz, and related … view at source ↗
Figure 10
Figure 10. Figure 10: Researchers are currently building multiple MEVIONs for large [PITH_FULL_IMAGE:figures/full_fig_p006_10.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

19 extracted references · 8 linked inside Pith

  1. [1]

    A Survey on Vision-Language-Action Models for Embodied AI,

    Y . Ma, Z. Song, Y . Zhuang, J. Hao, and I. King, “A Survey on Vision-Language-Action Models for Embodied AI,”arXiv preprint arXiv:2405.14093, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    A Survey on Vision-Language- Action Models: An Action Tokenization Perspective,

    Y . Zhong, F. Bai, S. Cai, X. Huang, Z. Chen, X. Zhang, Y . Wang, S. Guo, T. Guan, K. N. Luiet al., “A Survey on Vision-Language- Action Models: An Action Tokenization Perspective,”arXiv preprint arXiv:2507.01925, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Vision- Language-Action Models for Robotics: A Review Towards Real-World Applications,

    K. Kawaharazuka, J. Oh, J. Yamada, I. Posner, and Y . Zhu, “Vision- Language-Action Models for Robotics: A Review Towards Real-World Applications,”IEEE Access, vol. 13, pp. 162 467–162 504, 2025

    2025

  4. [4]

    Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system,

    Y . Qin, W. Yang, B. Huang, K. V . Wyk, H. Su, X. Wang, Y . Chao, and D. Fox, “Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system,”arXiv preprint arXiv:2307.04577, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Open-television: Teleoperation with immersive active visual feedback,

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang, “Open-television: Teleoperation with immersive active visual feedback,”arXiv preprint arXiv:2407.01512, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning,

    R. Ding, Y . Qin, J. Zhu, C. Jia, S. Yang, R. Yang, X. Qi, and X. Wang, “Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning,”arXiv preprint arXiv:2407.03162, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    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,” in2023 Robotics: Science and Systems, 2023

    2023

  8. [8]

    GELLO: A General, Low-cost, and Intuitive Teleoperation Framework for Robot Manip- ulators,

    P. Wu, Y . Shentu, Z. Yi, X. Lin, and P. Abbeel, “GELLO: A General, Low-cost, and Intuitive Teleoperation Framework for Robot Manip- ulators,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 12 156–12 163

    2024

  9. [9]

    Aloha 2: An enhanced low-cost hardware for bimanual teleoperation,

    J. Aldaco, T. Armstrong, R. Baruch, J. Bingham, S. Chan, K. Draper, D. Dwibedi, C. Finn, P. Florence, S. Goodrichet al., “Aloha 2: An enhanced low-cost hardware for bimanual teleoperation,”arXiv preprint arXiv:2405.02292, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,”arXiv preprint arXiv:2401.02117, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Alpha-αand bi-act are all you need: Importance of position and force information/control for imitation learning of unimanual and bimanual robotic manipulation with low-cost system,

    M. Kobayashi, T. Buamanee, and T. Kobayashi, “Alpha-αand bi-act are all you need: Importance of position and force information/control for imitation learning of unimanual and bimanual robotic manipulation with low-cost system,”IEEE Access, 2025

    2025

  12. [12]

    Face, “SO-101,” https://huggingface.co/docs/lerobot/so101, ac- cessed: 2026-07-05

    H. Face, “SO-101,” https://huggingface.co/docs/lerobot/so101, ac- cessed: 2026-07-05

    2026

  13. [13]

    OpenArm,

    Enactic, “OpenArm,” https://github.com/enactic/openarm, accessed: 2026-07-05

    2026

  14. [14]

    AhaRobot: A Low-Cost Open-Source Bimanual Mobile Manipulator for Embodied AI,

    H. Cui, Y . Yuan, Y . Zheng, and J. Hao, “AhaRobot: A Low-Cost Open-Source Bimanual Mobile Manipulator for Embodied AI,”arXiv preprint arXiv:2503.10070, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    MEVIUS: A Quadruped Robot Easily Constructed through E-Commerce with Sheet Metal Welding and Ma- chining,

    K. Kawaharazuka, S. Inoue, T. Suzuki, S. Yuzai, S. Sawaguchi, K. Okada, and M. Inaba, “MEVIUS: A Quadruped Robot Easily Constructed through E-Commerce with Sheet Metal Welding and Ma- chining,” in2024 IEEE-RAS International Conference on Humanoid Robots, 2024, pp. 631–636

    2024

  16. [16]

    MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception,

    K. Kawaharazuka, K. Yoneda, S. Inoue, T. Suzuki, J. Oda, and K. Okada, “MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception,”IEEE Robotics and Automation Practice, vol. 1, pp. 50–54, 2026

    2026

  17. [17]

    meviy (MISUMI),

    “meviy (MISUMI),” https://meviy.misumi-ec.com/worldwide/en/

  18. [18]

    Scikit-Robot,

    “Scikit-Robot,” https://github.com/iory/scikit-robot

  19. [19]

    The Pinocchio C++ library – A fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,

    J. Carpentier, G. Saurel, G. Buondonno, J. Mirabel, F. Lamiraux, O. Stasse, and N. Mansard, “The Pinocchio C++ library – A fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,” inIEEE International Symposium on System Integrations, 2019. DESCRIPTION OFSUPPLEMENTARYMATERIALS A. Overview The Supplementary Mate...

    2019