betaFeedback

WorksComputer Science

HandUMI: A Modular Hand-Worn Parallel-Jaw Interface for Robot-Free Bimanual Data Capture

Abstract

HandUMI captures robot-free bimanual demonstrations with a hand-worn parallel-jaw interface. Swappable fingertips match the robot's installed gripper; Quest/PICO controllers track hands and encoders measure aperture. Robot-agnostic recordings are retargeted by inverse kinematics to seven simulated bimanual arms (median position error ≤1.2\leq 1.2 mm; ≤4.7\leq 4.7% of episodes rejected). On bimanual towel folding, 562 robot-free demonstrations train policies on real PiPER arms: ACT succeeds in 12/20 attempts and fine-tuned π0.5π_{0.5} in 7/20; ACT on the first 101 and base π0.5π_{0.5} each succeed in 0/20. Hardware and software are open source (Apache-2.0), and a pair costs USD 557 with Quest 3S, excluding the optional ZED Mini and host computer. Project website: https://murobotics.ai/research/handumi.

Introduction

Robot teleoperation with leader arms, replicas or VR controllers [1, 2, 3, 4] requires robot access for each demonstration.

Robot-free interfaces such as UMI [5] decouple capture from arm kinematics, but contact-geometry or opening-motion mismatch may hinder transfer. YUBI [6], the closest, also tracks Quest controllers, measures aperture with an encoder and is open. HandUMI differs in three ways: the robot keeps its gripper instead of being refitted with YUBI’s; jaws translate, whereas YUBI records a jaw angle and models revolute fingers [6, 7]; and tips copy installed grippers, whereas its paper describes one finger design. We do not compare them directly. KIWI [8] offers interchangeable tool modules but estimates pose by offline SLAM and jaw width from video, with its software release only announced. FastUMI [9] adapts fingertips, and ActiveUMI [10] copies existing grippers in controller-held interfaces. Table 1 summarizes these choices.

DevicePose sourceEnc.BimanualOpen HWOpen SWTipsAudioUSD
UMI [5]mono-inertial SLAM×✓✓✓––742a
UMI-3D [12]LiDAR-inertial odom.××✓✓––1,400b
FastUMI [9]T265 VIO×✓–✓c✓–≈2,042d
DexUMI [13]iPhone ARKit✓–e✓✓n/a–≈2,448f
iPhUMI [14]iPhone ARKit×✓✓✓–✓g≈1,504h
RDT2 [15]Vive Tracker (IR)–✓–✓i––≈1,750j
HiFi-UMI [16]head SLAM + tags✓✓–––––
YUBI [6]Quest ctrl.✓✓✓✓––<750k
ActiveUMI [10]Quest ctrl.–✓––––>350l
EgoMI [17]Quest 3S ctrl.–✓◦◦––≈12,500m
XRZero-G0 [18]PICO 4 ctrl.–✓––n––>463o
KIWI [8]360° SLAM + IMU×✓✓◦✓p✓≈1,100q
HandUMIQuest / PICO ctrl.✓✓✓✓✓✓r557s

Table 1. Robot-free interfaces. Enc.: encoder-measured gripper state; Open HW/SW: public files/code; Tips: swappable tips. ✓/×: yes/no; ◦: announced; –: not established; n/a: not applicable. USD: named parts for a pair with tracking, excluding host, tax and shipping; >: only the headset is priced. a 2×(gripper USD 73 + GoPro kit USD 298), historical. b 2×(gripper USD 50 + LiDAR/camera USD 650). c Capture code and dataset. d 2×(GoPro 9 + T265) + HD60 X. e Unimanual evaluation. f 2×(iPhone 15 Pro Max + OAK-1 W + electronics); one hand published. g Contact microphone. h 2×(iPhone 15 Pro + UMI gripper). i Models and training/inference code. j 4 Vive Trackers + 2 Hikrobot cameras. k 2דunder USD 200” + Quest 3S (USD 349.99). l Quest 3S only. m 2×Robotiq 2F-85 + Quest 3S + ZED 2i. n Dataset only. o PICO 4 only. p Tool modules. q Two Insta360 X5; grippers unpriced. r Optional PICO audio, unused here. s USD 207.11 parts + Quest 3S; optional ZED Mini excluded.

HandUMI (Figure 1) contributes:

HandUMI operation setup, robot-free capture, gripper tracking, and multiple bimanual embodiments

Figure 1. HandUMI. Left: robot-free capture with the hand-worn device (fisheye camera, servo encoder, swappable fingertips) tracked by PICO 4 Ultra or Meta Quest controllers. Right: recorded trajectories retargeted in simulation to seven bimanual arms.

  • An open, low-cost hand-worn interface with translating parallel jaws, robot-matched swappable fingertips, live controller tracking and encoder-measured aperture; the robot retains its installed gripper.

  • A pipeline preserving robot-agnostic demonstrations and calibration provenance, with inverse kinematics (IK) and screening for seven bimanual arms.

  • Real-robot evidence on bimanual towel folding with PiPER arms: ACT [2] and fine-tuned π0.5\pi_{0.5} [11], trained only on retargeted HandUMI data, fold the towel in unassisted rollouts.

HandUMI

Hand-worn hardware. Modular mounts allow swapping fingertips, controller types and cameras on a common wearable body. Printed finger cradles drive parallel jaws through a crank (Figure 1). A Feetech STS3215 12-bit encoder measures jaw width via two-point calibration; an IMX335 fisheye records 1280×960 at 30 fps; an optional ZED Mini adds a workspace view. Fingertips match AgileX PiPER [19], ARX X5 [20], the Dream Gripper [21] for TRLC-DK1, Trossen WidowX AI [22] and the UMI soft finger [5]; the robot retains its gripper. Parts cost USD 207.11 per pair, or USD 557 with shared Quest 3S [23], excluding the optional ZED Mini, host computer, tax and shipping. All experiments used PICO 4 Ultra.

Tracking, synchronization and calibration. Instead of wrist-camera SLAM, a swappable provider supplies timestamped 6-DoF controller poses: Quest 3/3S through a new TCP app implementing the protocol of YUBI’s Apache-2.0 software [6], informed by analysis of its client APK, or PICO 4 Ultra through XRoboToolkit [24]. A ChArUco board [25, 26] defines the table frame. The workstation’s monotonic clock stamps every source; rows target 40 ms behind real time and take the nearest camera (30 fps) and encoder (100 Hz) samples with freshness flags (60 ms skew limit); sensor loss over 1 s discards the episode. Voice commands [27] or double-squeezes control recording. Gated calibration covers gripper range, intrinsics (≤0.8\le0.8 px mean reprojection error), camera mount and table frame (≤8\le8 mm RMS), and per-tip tool center point (TCP) pivot fits (<5<5 mm RMS, <10<10 mm maximum sample residual). A separate guided touch protocol checks both tips at one reference point (15 mm separation limit) and on the table (±15\pm15 mm height), using 20 samples per position by default; no result can be tied to the affected sessions. SHA-256 fingerprints identify configurations, not accuracy.

Source-preserving retargeting. LeRobot v3 recordings [28, 29] store two controller poses and jaw widths independently of the robot. Profiles supply kinematics, limits, TCPs and a linear gripper map (0–66 mm onto PiPER’s stroke; speed and force unmatched). PyRoki IK [30] generates joint data for PiPER, OpenArm v1 [31], TRLC-DK1 [32], I2RT YAM [33], MakerMods Metal Arm [34], Galaxea R1 Lite [35, 36] and ARX X5 [20]. Per-robot screening checks IK error, self-collision and base-joint swing. PiPER and OpenArm replay shares mapping, IK and smoothing with MuJoCo [37].

Experiments

We evaluate HandUMI on bimanual PiPER towel folding: the arms grasp two corners of a flat towel and fold it in half.

Capture quality

Dataset. Without a robot, three operators (authors; one with prior device experience) used PICO/PiPER-tip HandUMI to record the cohort over four days (about 5 h in total). Besides 3 unfolded towels and 1 tracking freeze, we dropped 79 of 641 episodes (12%) whose PiPER retargeting swings a base joint past 60°, biasing it toward PiPER’s workspace; the other arms accept 64–79 of these. The 562 kept episodes (201,397 frames, 111.9 min) pass the offline episode-quality gate: duration, finite values, tracking, sensor health, synchronization, pose jumps, freezes and aperture checks. Tracking, health and sync each allow up to 1% bad rows; health and sync exclude the first 0.2 s. A single row above 60 ms does not reject an episode. After faulty camera recalibration, 40 episodes received a post-hoc +1.9 cm height correction estimated from grasp contacts; 31 remain in the cohort. Fit residuals and offline checks do not independently verify physical table height; the reason the gates missed the fault is unverified.

Synchronization and calibration. Table 2 measures sampling-grid skew (≈8 ms median at 30 fps), not end-to-end latency. PICO leads ≈40 ms on the host clock; both wrists are tracked on 100.00% of rows. TCP fits (750 samples/side) pass both thresholds: RMS 4.0/3.4 mm; maximum sample-to-fitted-pivot Euclidean residuals 7.013/7.021 mm (left/right). These are internal consistency, not absolute accuracy. Table-frame RMS is 1.6–4.2 mm and 0.3–0.7°, two remounts differ 6–7 mm in TCP height.

StreamMed. (ms)p95 (ms)>60 ms (rows)
Left wrist cam.8.516.43
Right wrist cam.8.616.32
Top camera8.916.22
Encoder (100 Hz)2.54.84
PICO tracking40.340.40

Table 2. Selection skew over 198,025 rows (warm-up excluded).

Cross-embodiment retargeting. Table 3 replays 562 kept and 79 dropped episodes on seven simulated arms (Metal: only the 79, after a compute failure). PiPER uses the lab table; other placements are hyperparameters, not held out (OpenArm, ARX X5: swept on 19, 10 evaluated episodes). Rejected: unreachable start or p99 position error >3 cm. Self-collision: hull-confirmed episodes. Joint limit: frames at a limit. Up to 22.9% of episodes exceed URDF velocity limits (OpenArm). OpenArm is 7-DoF, others 6-DoF. PiPER-tip data test kinematics, not contact; PiPER’s rotation residual enters policy training, with unknown effect.

RobotPosition error (mm)Rejected (%)Self-collision (%)Joint limit (%)
PiPER1.170.00.00.1
OpenArm0.144.70.027.5
TRLC-DK10.183.12.72.3
YAM0.190.80.04.7
Metal0.090.00.00.4
R1 Lite0.030.016.40.0
ARX X50.090.00.00.1

Table 3. Retargeting of 641 episodes (Metal: 79). Position error is the median.

Robot-free data for policies

Setup. Policies use HandUMI IK commands as 14-D PiPER states/actions (radians, grippers in [0, 1]) and two wrist views, excluding the top view (Figure 2). PiPER retains the same IMX335 cameras and gripper-relative positions. Real PiPER open-loop replay validated recorded trajectories. ACT [2] uses one recipe for 101 episodes (19.3 min) and the full cohort; π₀.₅ [11] uses full-cohort fine-tuning and a base-model reference. A 10% episode holdout selects the lowest-validation-loss checkpoint (ACT: batch 96, lr 10−510^{-5}, chunk 100; π₀.₅: batch 16, lr 4×10−54 \times10^{-5}, chunk 50).

Figure 2. Towel folding: views, a rollout, and successes by training episodes (Wilson 95% CI).

Protocol. Each model runs 20 unassisted attempts at random towel positions within the training area, unmatched across models. The operator judges a final fold in half live, without numeric tolerance. Videos: project website; data and checkpoints: https://huggingface.co/murobotics.

Results. ACT on all 562 episodes succeeds in 12/20 attempts (95% Wilson CI 39–78%) and fine-tuned π₀.₅ in 7/20 (18–57%; Figure 2); at n=20n=20 they are not statistically distinguishable (Fisher exact p=0.20p=0.20). ACT on the first 101 episodes and π₀.₅ without fine-tuning each succeed in 0/20 (0–16%). The former is a scale probe, confounded by added sessions/calibrations; the latter is a sanity check for HandUMI fine-tuning.

Conclusion and Limitations

HandUMI abstracts arm kinematics while retaining the robot’s gripper; new grippers require new capture tips. Contact geometry remains embodiment-dependent. Robot-free data enable real PiPER towel folding (ACT 12/20, π₀.₅ 7/20).

Limitations. Cross-embodiment evidence is simulated IK; real tests cover one PiPER task, setup and seed, without failure analysis; holdouts can share sessions. Capture is headset-bounded and parallel-jaw; tip-matching gain, jaw accuracy, fingertip load and audio value are unmeasured.

Acknowledgments

We thank the NONHUMAN community for their collaboration on data collection and device testing.

References

  1. [1]A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, et al. DROID: A large-scale in-the-wild robot manipulation dataset. In Robotics: Science and Systems (RSS), 2024. doi:10.15607/RSS.2024.XX.120. URL https://www.roboticsproceedings.org/rss20/p120.html.DOI
  2. [2]T. Z. Zhao, V. Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In Robotics: Science and Systems XIX, 2023. doi:10.15607/RSS.2023.XIX.016. URL https://tonyzhao28.github.io/aloha/.DOI
  3. [3]Z. Fu, T. Z. Zhao, and C. Finn. Mobile ALOHA: Learning bimanual mobile manipulation using low-cost whole-body teleoperation. In Proceedings of The 8th Conference on Robot Learning, volume 270 of Proceedings of Machine Learning Research, pages 4066–4083. PMLR, 2025. URL https://proceedings.mlr.press/v270/fu25b.html.arxiv.org/abs/2401.02117
  4. [4]P. Wu, Y. Shentu, Z. Yi, X. Lin, and P. Abbeel. GELLO: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163, 2024. doi:10.1109/IROS58592.2024.10801581. URL https://doi.org/10.1109/IROS58592.2024.10801581.
  5. [5]C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. In Robotics: Science and Systems, 2024. doi:10.15607/RSS.2024.XX.045. URL https://www.roboticsproceedings.org/rss20/p045.html.DOI
  6. [6]T. Ohkawa, J. Arima, Y. Noguchi, M. Tateno, M. Sugiura, T. Okubo, K. Ikeuchi, Y. Shin, H. Nishizawa, N. Kanazawa, Y. Wakayama, D. Fukunaga, K. Makihara, T. Motoda, F. Erich, Y. Domae, T. Matsushima, Y. Okumatsu, and K. Ota. YUBI: Yielding universal bidigital interface for bimanual dexterous manipulation at scale. arXiv preprint arXiv:2606.10244, 2026. URL https://arxiv.org/abs/2606.10244.
  7. [7]YUBI Contributors. YUBI hand description (URDF/xacro). GitHub repository, 2026. URL https://github.com/airoa-org/yubi-sw/blob/1eccb041efc4d8966b03e4ccf4b402aa79d78be4/yubi_description/urdf/yubi_hand.urdf.xacro. Revision 1eccb04; accessed 2026-10-15.
  8. [8]B. C. Yang, W. Wang, S. Li, K. Yan, S. Wilkinson, Z. Wang, Y. F. Yeung, and L. Sun. Kinematic interface for the wild: Modular bimanual loco-manipulation capture from 360° cameras alone. arXiv preprint arXiv:2609.22809, 2026. URL https://arxiv.org/abs/2609.22809.
  9. [9]Zhaxizhuoma, K. Liu, C. Guan, Z. Jia, Z. Wu, X. Liu, T. Wang, S. Liang, P. Chen, P. Zhang, H. Song, D. Qu, D. Wang, Z. Wang, N. Cao, Y. Ding, B. Zhao, and X. Li. FastUMI: A scalable and hardware-independent universal manipulation interface with dataset. In Proceedings of The 9th Conference on Robot Learning, volume 305 of Proceedings of Machine Learning Research, pages 3069–3093. PMLR, 2025. URL https://proceedings.mlr.press/v305/zhaxizhuoma25a.html.arxiv.org/abs/2409.19499
  10. [10]Q. Zeng, C. Li, J. St. John, Z. Zhou, J. Wen, Y. Xu, G. Feng, and Y. Zhu. ActiveUMI: Robotic manipulation with active perception from robot-free human demonstrations. In IEEE International Conference on Robotics and Automation (ICRA), pages 14273–14280, 2026. doi:10.1109/ICRA57385.2026.11696372. URL https://ieeexplore.ieee.org/document/11696372.DOI
  11. [11]K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y. Galliker, D. Ghosh, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. Vuong, H. Walke, A. Walling, H. Wang, L. Yu, and U. Zhilinsky. π0.5: A vision-language-action model with open-world generalization. In Proceedings of The 9th Conference on Robot Learning, volume 305 of Proceedings of Machine Learning Research, pages 17–40. PMLR, 2025. URL https://proceedings.mlr.press/v305/black25a.html.arxiv.org/abs/2504.16054
  12. [12]Z. Wang. UMI-3D: Extending universal manipulation interface from vision-limited to 3D spatial perception. arXiv preprint arXiv:2604.14089, 2026. URL https://arxiv.org/abs/2604.14089.
  13. [13]M. Xu, H. Zhang, Y. Hou, Z. Xu, L. Fan, M. Veloso, and S. Song. DexUMI: Using human hand as the universal manipulation interface for dexterous manipulation. In Proceedings of The 9th Conference on Robot Learning, volume 305 of Proceedings of Machine Learning Research, pages 437–459. PMLR, 2025. URL https://proceedings.mlr.press/v305/xu25b.html.arxiv.org/abs/2505.21864
  14. [14]A. Patel, B. Pekarek, J. E. Castro Hernandez, and S. Song. What enables in-context behavior prompting for manipulation? arXiv preprint arXiv:2606.30457, 2026. URL https://arxiv.org/abs/2606.30457.
  15. [15]S. Liu, B. Li, K. Ma, L. Wu, H. Tan, X. Ouyang, H. Su, and J. Zhu. RDT2: Exploring the scaling limit of UMI data towards zero-shot cross-embodiment generalization. arXiv preprint arXiv:2602.03310, 2026. URL https://arxiv.org/abs/2602.03310.
  16. [16]Simple AI, Y. Wei, J. Ma, J. Wang, W. Zhou, Y. Zuo, K. Rui, M. Li, J. Zhang, Z. Pan, X. Wang, H. Jia, H. Du, Z. Zeng, J. Ma, G. Qin, D. Zhang, and X. Li. HiFi-UMI: Learning deployable manipulation policies from high-fidelity UMI data alone. arXiv preprint arXiv:2607.25895, 2026. URL https://arxiv.org/abs/2607.25895.
  17. [17]J. Yu, Y. Shentu, D. Wu, P. Abbeel, K. Goldberg, and P. Wu. EgoMI: Learning active vision and whole-body manipulation from egocentric human demonstrations. arXiv preprint arXiv:2511.00153, 2025. URL https://arxiv.org/abs/2511.00153.
  18. [18]J. Wang, P. Pu, Z. Fung, A. Wang, S. Wang, B. Deng, K. Wang, Z. Liu, C. Pan, P. Yang, A. Zhai, L. Liang, S. Li, J. Sun, J. Xu, W. Tian, K. Yan, K. Ye, S. Li, Q. Wang, R. Gan, and H. Wang. XRZero-GO: Pushing the frontier of dexterous robotic manipulation with interfaces, quality and ratios. arXiv preprint arXiv:2604.13001, 2026. URL https://arxiv.org/abs/2604.13001.
  19. [19]AgileX Robotics. PiPER: 6-DoF robotic arm. https://global.agilex.ai/products/piper, 2024. SDK: https://github.com/agilexrobotics/piper_sdk. Accessed 2026-10-16.
  20. [20]ARX Robotics. ARX X5 robotic arm software. https://github.com/ARXroboticsX/ARX_X5, 2025. Accessed 2026-10-16.
  21. [21]Dream Machines Robotics. Dream Gripper: Open-source 3D-printable grippers for the TRLC-DK1. https://github.com/Dream-Machines-Robotics/Dream-Gripper, 2026. Accessed 2026-10-16.
  22. [22]Trossen Robotics. WidowX AI robotic arm. https://www.trossenrobotics.com/widowx-ai, 2026. Accessed 2026-10-16.
  23. [23]Meta. An update on Meta Quest pricing. https://www.meta.com/blog/update-meta-quest-pricing/, 2026. Published 2026-04-16; US prices effective 2026-04-19. Accessed 2026-10-10.
  24. [24]Z. Zhao, L. Yu, K. Jing, and N. Yang. XRoboToolkit: A cross-platform framework for robot teleoperation. In IEEE/SICE International Symposium on System Integration (SII), pages 15–20, 2026. doi:10.1109/SII64115.2026.11404528. URL https://arxiv.org/abs/2508.00097.
  25. [25]OpenCV. Detection of ChArUco boards. https://docs.opencv.org/4.13.0/d4/da/tutorial_charuco_detection.html, 2025. OpenCV 4.13.0 documentation. Accessed 2026-10-10.
  26. [26]S. Garrido-Jurado, R. Muñoz-Salinas, F. J. Madrid-Cuevas, and M. J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280–2292, 2014. doi:10.1016/j.patcog.2014.01.005.
  27. [27]Alpha Cephei. Vosk: Offline speech recognition API. https://github.com/alphacep/vosk-api, 2019. Model vosk-model-small-en-us-0.15, https://alphacephei.com/vosk/models. Accessed 2026-10-10.
  28. [28]R. Cadene, S. Alibert, A. Soare, Q. Gallouédec, A. Zouitine, S. Palma, P. Kooijmans, M. Aractingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, K. Mեftah, M. Ellerbach, J. Moss, and T. Wolf. LeRobot: State-of-the-art machine learning for real-world robotics in PyTorch. https://github.com/huggingface/lerobot, 2024.
  29. [29]R. Cadene, S. Alibert, F. Capuano, M. Aractingi, A. Zouitine, P. Kooijmans, J. Choghari, M. Russi, C. Pascal, S. Palma, D. Aubakirova, M. Shukor, J. Moss, A. Soare, Q. Lhoest, Q. Gallouédec, and T. Wolf. LeRobot: An open-source library for end-to-end robot learning. In International Conference on Learning Representations (ICLR), 2026. URL https://openreview.net/forum?id=CiZMMAFQR3. arXiv:2602.22818.arxiv.org/abs/2602.22818
  30. [30]C. M. Kim, B. Yi, H. Choi, Y. Ma, K. Goldberg, and A. Kanazawa. PyRoki: A modular toolkit for robot kinematic optimization. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1312–1319, 2025. doi:10.1109/IROS60139.2025.11246651. URL https://arxiv.org/abs/2505.03728.
  31. [31]Enactic, Inc. OpenArm: A fully open-source humanoid arm for physical AI research and deployment in contact-rich environments. https://github.com/enactic/openarm, 2025. Documentation (v1.0): https://docs.openarm.dev/1.0/. Accessed 2026-10-16.
  32. [32]The Robot Learning Company. TRLC-DK1: An open source dev kit for AI-native robotics. https://github.com/robot-learning-co/trlc-dk1, 2025. Accessed 2026-10-16.
  33. [33]I2RT Robotics. YAM robotic arm. https://i2rt.com/products/yam-6-dof-arm, 2025. SDK: https://github.com/i2rt-robotics/i2rt. Accessed 2026-10-16.
  34. [34]MakerMods. Metal Arm: 6-axis aluminum robot arm. https://www.makermods.ai/metal-arm, 2026. SDK: https://github.com/makermods-robotics/metal-python-ros. Accessed 2026-10-16.
  35. [35]Galaxea Dynamics. Galaxea R1 Lite: General mobile manipulation platform. https://galaxea-dynamics.com/products/galaxea-r1-lite-6-dof-general-mobile-manipulation-platform, 2026. Accessed 2026-10-16.
  36. [36]T. Jiang, T. Yuan, Y. Liu, C. Lu, J. Cui, X. Liu, S. Cheng, J. Gao, H. Xu, and H. Zhao. Galaxea open-world dataset and G0 dual-system VLA model. arXiv:2509.00576, 2025. URL https://arxiv.org/abs/2509.00576.
  37. [37]E. Todorov, T. Erez, and Y. Tassa. MuJoCo: A physics engine for model-based control. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5026–5033, 2012. doi:10.1109/IROS.2012.6386109.DOI

Paper details

vlnc:W_yy4eptvt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents