Pith. sign in

REVIEW 4 major objections 6 minor 35 references

A single VR operator can teleoperate a miniature humanoid's arms and legs simultaneously, with a reinforcement-learned balance policy that transfers to hardware without retuning.

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 09:55 UTC pith:KIUYLNCZ

load-bearing objection A genuine feasibility demo on miniature humanoid tele-loco-manipulation, but the headline speed claim is unmeasured and the evaluation is thin; still worth a serious referee. the 4 major comments →

arxiv 2607.20399 v1 pith:KIUYLNCZ submitted 2026-07-22 cs.RO cs.HCcs.LG

Towards Miniature Humanoid Tele-Loco-Manipulation Using Virtual Reality and Reinforcement Learning

Nicolas Kosanovic , Jordan Dowdy , Jean Chagas Vaz This is my paper
classification cs.RO cs.HCcs.LG
keywords humanoid teleoperationvirtual realityreinforcement learninglocomotionsim-to-real transfertorque controlDYNAMIXELtele-loco-manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the powerful combination of virtual-reality teleoperation and reinforcement-learned locomotion, normally reserved for full-sized humanoids, can be built from the ground up for miniature robots. The authors construct a full-stack system for DYNAMIXEL-based humanoids, demonstrated on the ROBOTIS OP3, where an operator wearing a VR headset controls the arms through inverse kinematics and the legs through a blind RL walking policy. They report walking speeds up to 0.45 m/s independent of arm motion, and a cube-relocation experiment where the teleoperated robot moved two 40 g cubes over roughly 5 m in 10 minutes. The result matters because it opens miniature humanoids to telepresence research and egocentric demonstration data collection at a fraction of the cost of full-sized platforms.

Core claim

The central claim is that a miniature humanoid can be given full-body tele-loco-manipulation using VR for the upper body and RL for lower-body balance, with zero-shot sim-to-real transfer of the locomotion policy. The authors achieve this by writing all robot software from scratch, including a 200 Hz torque controller that uses hardware-level current control on the DYNAMIXEL XM430 servos, and by identifying actuator dynamics through a pendulum test bench (armature 0.045, friction loss 0.03, effort limit 3.6). The trained policy uses domain randomization on link mass, gravity, friction, and body perturbations, plus an observation history, to handle the mass shifts caused by arm motion. Experi

What carries the argument

The load-bearing mechanism is the pairing of a data-driven actuator model with a torque-based impedance controller: accurate servo dynamics identified from pendulum test-bench data enable the reinforcement-learned locomotion policy to transfer zero-shot from simulation to real hardware, while a 200 Hz PD torque controller gives the arms compliant, backdrivable behavior. The policy's observation space—proprioceptive data with a nine-step history—lets it infer joint velocities and reject body perturbations without explicit state estimation.

Load-bearing premise

The entire result depends on the RL balance policy transferring from simulation to real hardware without retuning, which requires the actuator dynamics model to be accurate enough and the domain randomization broad enough to cover the mass and center-of-mass shifts caused by arm motion.

What would settle it

Measure the robot's actual walking speed on hardware while the arms move; if it cannot sustain 0.45 m/s or falls when arm configurations shift the center of mass beyond the randomized range, the independent-arm-motion/balance claim collapses.

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

If this is right

  • A single operator can command arm and leg motions independently in real time, with the RL balance policy absorbing arm-induced mass shifts without retuning.
  • The reported 0.45 m/s walking speed is fast for a miniature humanoid and is maintained independently of arm motion.
  • Because the torque controller is robot-agnostic for DYNAMIXEL actuator chains, the same stack can be ported to other miniature humanoid morphologies.
  • The impedance-controlled arms permit compliant contact, reducing the risk of damage to the robot or its environment.
  • The teleoperation stack can collect egocentric demonstrations on miniature hardware, which the authors argue could improve imitation learning data.

Where Pith is reading between the lines

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

  • The paper's 0.45 m/s speed figure is not backed by a hardware speed measurement; the locomotion validation was an 80-second random-command trial, so the actual sustained walking speed during teleoperation remains to be measured directly.
  • Replacing the touchpad velocity commands with the planned VR treadmill interface could make locomotion more intuitive and perhaps improve the modest 2-of-6 cube relocation rate.
  • If the actuator model and randomization suffice, the stack should transfer to other DYNAMIXEL humanoids, enabling cross-platform comparisons of tele-loco-manipulation.
  • The force estimate via virtual/translucent model separation is adequate for slow contacts; hard or dynamic contacts would likely require real force sensing, a limitation consistent with the paper's note that the system was 'too soft' and required crossed arms to hold cubes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper presents an integrated tele-loco-manipulation stack for a miniature ROBOTIS OP3 humanoid. Upper-body VR teleoperation uses VIVE tracking, a Unity-based interface, inverse kinematics, and a custom open-loop current-based torque/impedance controller for the DYNAMIXEL arms. Lower-body balance and locomotion are handled by an RL policy trained in Isaac Lab/Isaac Sim with domain randomization and BAM-based actuator parameters, deployed zero-shot. Hardware experiments measure upper-body tracking latency (~220 ms), an 80 s random-velocity locomotion trial that keeps projected gravity near -1, and a 10-minute bimanual cube-relocation task with average 2/6 blocks moved and ~0.35 m/s walking speed. The abstract additionally claims walking up to 0.45 m/s independent of arm motions.

Significance. The principal contribution is architectural: a complete, low-cost, compliant VR-to-robot control stack for a small humanoid, including actuator system identification, PD torque control, and RL locomotion. The hardware experiments and honest discussion of limitations are useful, and the BAM-based actuator identification is a concrete step toward reproducible sim-to-real transfer. However, the headline performance claim (0.45 m/s independent of arm motions) is not substantiated by any reported measurement, and several quantitative experiments are reported with insufficient detail. If the missing speed and robustness data can be supplied, the framework would be a valuable reference for miniature-humanoid tele-loco-manipulation; in its current form the results are preliminary.

major comments (4)
  1. [Abstract; §V.A.3] The abstract's central claim of "walking at speeds up to 0.45 m/s independent of arm motions" has no corresponding measurement. Section IV.B sends random velocity commands for 80 s but reports only normalized projected gravity, not forward speed, distance, step count, or falls. The only measured speed is the ~0.35 m/s average in the tele-loco experiment (§V.A.3), which involved arm motions, a tethered power supply, and only 2/6 block relocation. Please add a dedicated speed-validation trial with commanded vs. actual velocity (e.g., motion capture or odometry), report the maximum sustained speed actually achieved, and state whether arm motions were present. Without this, the 0.45 m/s figure overstates the demonstration.
  2. [§IV.B; §V.A.2] The locomotion robustness evaluation appears to be a single 80-second run with no quantitative outcome criterion. To support the claim that balance is robust to randomly commanded whole-body velocities and to operator arm motions, the authors should report the number of repeated trials, the number of falls/stumbles, and a quantitative measure of speed-tracking error or recovery behavior. Projected gravity staying near -1 is necessary but not sufficient evidence: a robot could be upright yet fail to track the commanded velocity or rely on the tether. Please provide per-trial outcomes and statistics.
  3. [§V.A.3; Abstract] The tele-loco-manipulation result is reported ambiguously and without variance: Abstract says "moved 2 different 40g cubes", while §V.A.3 says "2 out of 6 blocks"; "Ten trials are completed" but only a single average is given. It is unclear whether 2/6 is the mean number of relocated blocks over ten trials, and there is no standard deviation, range, or trial-to-trial table. Please clarify the task definition (six available cubes? two required?), report the full distribution of outcomes, and reconcile the abstract with the results. This is important because the system's overall utility is judged on this task.
  4. [§III.C.2] The zero-shot sim-to-real claim rests on the BAM actuator-model parameters (armature 0.045, friction loss 0.03, effort limit 3.6), but no quantitative fit error between the pendulum test-bed data and the simulated actuator response is reported. Figure 3 is qualitative. Please include numerical agreement metrics (e.g., RMSE in joint angle/current for step and sine-sweep tests) and state how sensitive the final locomotion policy was to those parameters. Without this, the reader cannot assess how load-bearing the actuator model fidelity actually is, especially given Section VI calls the model "rudimentary".
minor comments (6)
  1. [§III.B.3, Eq. (2)] The condition "if ||x|| < 0" is impossible as written; it should be "if ||x|| < x_threshold". Please also clarify the sign/direction convention of the force vector and state whether F is applied to the virtual or real robot.
  2. [§III.C.1, Eqs. (4)-(5)] Equation (5) is the definition of a Q-value, not a policy objective; Equation (4) already defines π*. Please correct or remove the duplicate/incorrect expression.
  3. [Table IV] Only right-arm tracking errors are reported. Please include the left-arm values or state explicitly that the two arms are assumed symmetric and only one was analyzed.
  4. [§V.B] Typo: "RL-basd" should be "RL-based".
  5. [References] Reference [6] misspells the author name as "Sciliano"; should be "Siciliano".
  6. [§III.B.1] The phrase "less than a 4 M bps bitrate" is awkward; consider "less than 4 Mbps".

Circularity Check

0 steps flagged

No circularity: the locomotion policy is trained in simulation and transferred to hardware; self-citations are contextual, not load-bearing.

full rationale

Walking the derivation chain, no load-bearing step reduces to its own input. The RL locomotion policy is trained in Isaac Sim with domain randomization and BAM actuator parameters fitted from a pendulum test bench; those parameters are inputs to the simulation, and the hardware trial is a forward zero-shot sim-to-real transfer, not a prediction of the fitted target. The tele-loco result (2 cubes in 10 minutes at about 0.35 m/s average) is a measured outcome of a hardware experiment. The abstract's 0.45 m/s figure is not tied to a reported measurement, and Section VI concedes that the actuator model is 'rudimentary' and demands 'excessive domain randomization'; that is a validation gap and a correctness risk, not circularity. Equations (1)-(5) are standard PD-torque and MDP definitions and do not introduce any target quantity by construction. Self-citations ([5], [13], [25], [34]) provide context for embodiment scaling, VR UI, and domain randomization; none is invoked as a uniqueness theorem or as the source of the locomotion/balance result. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper's empirical claims rest on several manually chosen control gains and reward weights plus a fitted actuator model, but no new physical entities, forces, or conserved quantities are introduced. The Hooke's-law force estimate in Eq. (2) is a display/feedback heuristic, not a new entity.

free parameters (6)
  • PD gains K_P, K_D (Eq. 1)
    Manual joint-stiffness and damping gains for 20 joints; lower on arms/head for compliance; values not reported; central to the torque controller in Section III-B2.
  • BAM actuator parameters (armature, friction loss, effort limit) = 0.045, 0.03, 3.6
    Fitted from pendulum test-bench data in Section III-C2 and used to set the actuator model in Isaac Sim for zero-shot policy transfer.
  • Domain-randomization ranges (Table I)
    Hand-chosen ranges for foot friction, gravity, mass, body velocity/position/orientation, joint states; selected to bridge sim2real and not derived from data.
  • Reward and penalty weights (Table III)
    Hand-tuned weights for seven rewards and ten penalties; these shape the learned walking policy and are central to the claimed stability.
  • Observation noise injection magnitudes (Table II)
    Hand-chosen noise values added to observations during training; affect policy robustness and are not fitted to data.
  • Haptic force-estimation stiffness K and threshold x_threshold (Eq. 2)
    Chosen static gains for the virtual/translucent-model distance-based force proxy; no calibration against real contact forces is reported.
axioms (5)
  • domain assumption DYNAMIXEL XM430 output torque is proportional to commanded current, with motor constant from stall torque/current.
    Section III-B2: underlies every torque command in Eq. (1); supported only by a preliminary static testbed, with no load or temperature characterization.
  • domain assumption The BAM pendulum-bench actuator model plus domain randomization is sufficient for zero-shot sim-to-real on OP3.
    Section III-C2: the locomotion policy is never fine-tuned on hardware, so the whole lower-body claim depends on model fidelity and randomization coverage.
  • domain assumption Arm motions can be treated as mass/platform disturbances to the RL balance policy, without a whole-body dynamic coupling model.
    Section III-B2 and III-C2: the system separates upper-body IK and lower-body RL; no coupling dynamics are modeled beyond randomized perturbations during training.
  • domain assumption Normalized projected gravity near -1 is a valid indicator of bipedal stability.
    Section IV-B and Figure 6: used as the sole quantitative stability metric, with no comparison to ground-truth CoM/ZMP or fall-rate data.
  • domain assumption The RL training setup in Isaac Lab is adequate and converges to a policy that transfers; exact hyperparameters are not reported.
    Section III-C: observations, rewards, and randomization are given, but network architecture, PPO hyperparameters, and training time are omitted, so reproducibility of the learned policy depends on unspecified choices.

pith-pipeline@v1.3.0-alltime-deepseek · 10337 in / 12623 out tokens · 107636 ms · 2026-08-01T09:55:37.568209+00:00 · methodology

0 comments
read the original abstract

Full-sized humanoid robot capabilities have grown exponentially in recent years, aiming towards general-purpose deployment in human environments. A popular control method used by manufacturers utilizes Virtual Reality for upper-body teleoperation and Reinforcement Learning for lower-body balance and locomotion control. As a result, a single remote operator can see, manipulate, and navigate about a real, distant physical environment. This powerful control stack is often relegated to expensive full-sized robots, many of which are inaccessible to the research community. Miniature humanoids are more prevalent, but employ less biomimicry in their design (e.g. fewer sensors, Degrees of Freedom, etc) and lack similar developments. This paper describes a compliant full-body telepresence control stack developed from the ground up for miniature humanoids. Framework experimentation on ROBOTIS OP3 hardware showcases walking at speeds up to 0.45 m/s independent of arm motions. Tele-loco-manipulation is demonstrated via a cube relocation experiment with an expert human operator. On average, the teleoperated system moved 2 different 40 g cubes within 10 mins, walking a total distance of 5 m. Overall, the developed system shows potential for miniature humanoid tele-loco-manipulation.

Figures

Figures reproduced from arXiv: 2607.20399 by Jean Chagas Vaz, Jordan Dowdy, Nicolas Kosanovic.

Figure 1
Figure 1. Figure 1: Human teleoperator using the proposed tele-loco-manipulation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ”Operator view” of the remote environment from images streamed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Real2Sim PD gain tuning for the robot’s actuators (Dynamixel [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results from telemanipulation experiments, comparing human hand to robot end effector motion. The plots on the left show transient data in the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Snapshots from a real tele-loco-manipulation experiment trial that visually depict the act of block relocation. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Humanoid robot’s projected gravity during random velocity [PITH_FULL_IMAGE:figures/full_fig_p007_6.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

35 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Introducing NEO Gamma,

    1X Technologies, “Introducing NEO Gamma,”1X Technologies, Feb. 21, 2025. [Online]. Available: https://www.1x.tech/discover/introducing-neo-gamma

    2025

  2. [2]

    [Online]

    Unitree Robotics, ”Unitree G1 Humanoid”,Unitree Robotics. [Online]. Available: https://www.unitree.com/g1/

  3. [3]

    Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning

    Z. Guet al., “Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning”, 2025, arXiv:2501.02116

    Pith/arXiv arXiv 2025

  4. [4]

    Open X-Embodiment: Robotic Learning Datasets and RT-X Models : Open X-Embodiment Collaboration,

    A. O’Neillet al., “Open X-Embodiment: Robotic Learning Datasets and RT-X Models : Open X-Embodiment Collaboration,” 2024 IEEE International Conference on Robotics and Automation (ICRA), Yokohama, Japan, 2024, pp. 6892-6903, doi: 10.1109/ICRA57147.2024.10611477

    arXiv 2024

  5. [5]

    Avatar-Darwin a Social Humanoid with Telepresence Abilities Aimed at Embodied Avatar Systems,

    A. Dave, J. C. Vaz, J. Kim, N. Kosanovic, N. Kassai and P. Y . Oh, “Avatar-Darwin a Social Humanoid with Telepresence Abilities Aimed at Embodied Avatar Systems,”2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), Ginowan, Japan, 2022, pp. 47-52, doi: 10.1109/Humanoids53995.2022.10000176

    arXiv 2022

  6. [6]

    Sciliano and O

    B. Sciliano and O. Khatib,Springer Handbook of Robotics, 1st ed., Heidelberg, Springer Berlin, 2008, pp. 719-739. [Online]. Available: https://doi.org/10.1007/978-3-540-30301-5

    doi:10.1007/978-3-540-30301-5 2008

  7. [7]

    Expanding Humanoid’s Material-Handling Capabilities using Capture Point Walking,

    J. C. Vaz and P. Y . Oh, “Expanding Humanoid’s Material-Handling Capabilities using Capture Point Walking,”2020 American Control Conference (ACC), Denver, CO, USA, 2020, pp. 2082-2087, doi: 10.23919/ACC45564.2020.9147541

    arXiv 2020

  8. [8]

    Model-Based Suppression Control for Liquid Vessels Carried by a Humanoid Robot While Stair-Climbing,

    J. C. Vaz and P. Oh, “Model-Based Suppression Control for Liquid Vessels Carried by a Humanoid Robot While Stair-Climbing,”2020 IEEE 16th International Conference on Automation Science and Engineering (CASE), Hong Kong, China, 2020, pp. 1540-1545, doi: 10.1109/CASE48305.2020.9216826

    arXiv 2020

  9. [9]

    Analysis and Perspectives on the ANA Avatar XPRIZE Competition

    K. Hauseret al., “Analysis and Perspectives on the ANA Avatar XPRIZE Competition”, inIntl. Jour. of Soc. Robotics, Jan. 2024. Accessed: Feb. 19, 2025. doi: 10.48550/arXiv.2401.05290 [Online]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.05290 2024

  10. [10]

    Remote telepresence over large distances via robot avatars: case studies,

    M. Elobaidet al., “Remote telepresence over large distances via robot avatars: case studies,”2024 IEEE Conference on Telepresence, Pasadena, CA, USA, 2024, pp. 183-187, doi: 10.1109/Telepresence63209.2024.10841750

    arXiv 2024

  11. [11]

    iCub3 avatar system: Enabling remote fully immersive embodiment of humanoid robots,

    S. Dafarraet al., “iCub3 avatar system: Enabling remote fully immersive embodiment of humanoid robots,” inSci. Robot., vol. 9, eadh3834, 2024, doi: 10.1126/scirobotics.adh3834

    doi:10.1126/scirobotics.adh3834 2024

  12. [12]

    Approach of Team Avatar-Hubo to the ANA Avatar XPRIZE Finals,

    B. Kim, N. Kassai, Z. Castrejon, N. Kosanovic, J. C. Vaz and P. Oh, “Approach of Team Avatar-Hubo to the ANA Avatar XPRIZE Finals,”2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN), Pasadena, CA, USA, 2024, pp. 37-42, doi: 10.1109/RO-MAN60168.2024.10731240

    arXiv 2024

  13. [13]

    ART: Avatar Robotics Telepresence— the future of humanoid material handling loco-manipulation,

    J. Vaz, N. Kosanovic and P. Oh, “ART: Avatar Robotics Telepresence— the future of humanoid material handling loco-manipulation,” inIntel Serv Robotics, Dec. 2023. Accessed: Feb. 2025. doi: 10.1007/s11370- 023-00499-x [Online]

    work page doi:10.1007/s11370- 2023

  14. [14]

    Humanoids in Hospitals: A Technical Study of Humanoid Surrogates for Dexterous Medical Interventions

    S. Ataret al., “Humanoids in Hospitals: A Technical Study of Humanoid Surrogates for Dexterous Medical Interventions”, 2025, arXiv:2503.12725

    Pith/arXiv arXiv 2025

  15. [15]

    High-Speed and Impact Resilient Teleoperation of Humanoid Robots,

    S. Bertrandet al., “High-Speed and Impact Resilient Teleoperation of Humanoid Robots,” Sept. 2024, arXiv:2409.04639 [cs.RO]

    Pith/arXiv arXiv 2024

  16. [16]

    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”, 2024, arXiv:2407.01512

    Pith/arXiv arXiv 2024

  17. [17]

    Deep Imitation Learning for Humanoid Loco-manipulation Through Human Teleoperation,

    M. Seoet al., “Deep Imitation Learning for Humanoid Loco-manipulation Through Human Teleoperation,”2023 IEEE-RAS 22nd International Conference on Humanoid Robots (Humanoids), Austin, TX, USA, 2023, pp. 1-8, doi: 10.1109/Humanoids57100.2023.10375203

    arXiv 2023

  18. [18]

    HOMIE Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit

    Q. Benet al., “HOMIE Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit”, 2025 Robotics: Science and Systems, 2025, arXiv: arXiv:2502.13013

    Pith/arXiv arXiv 2025

  19. [19]

    Mobile-TeleVision: Predictive Motion Priors for Humanoid Whole-Body Control

    C. Luet al., “Mobile-TeleVision: Predictive Motion Priors for Humanoid Whole-Body Control”, 2025, arXiv:2412.07773

    Pith/arXiv arXiv 2025

  20. [20]

    Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning

    N. Rudin, D. Hoeller, P. Reist and M. Hutter, “Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning”,2021 Conference on Robot Learning (CoRL), 2021, doi: arXiv:2109.11978

    Pith/arXiv arXiv 2021

  21. [21]

    Sim-to-Real Transfer of Compliant Bipedal Locomotion on Torque Sensor-Less Gear-Driven Humanoid

    S. Masuda and K. Takahashi, “Sim-to-Real Transfer of Compliant Bipedal Locomotion on Torque Sensor-Less Gear-Driven Humanoid”, 2025, arXiv:2204.03897

    Pith/arXiv arXiv 2025

  22. [22]

    Learning Bipedal Locomotion on Gear-Driven Humanoid Robot Using Foot- Mounted IMUs

    S. Katayama, Y . Koda, N. Nagatsuka and M. Kinoshita “Learning Bipedal Locomotion on Gear-Driven Humanoid Robot Using Foot- Mounted IMUs”, 2025, arXiv:2504.00614

    Pith/arXiv arXiv 2025

  23. [23]

    Learning agile soccer skills for a bipedal robot with deep reinforcement learning

    T. Haarnojaet al., “Learning agile soccer skills for a bipedal robot with deep reinforcement learning”,Science Robotics, vol. 9, Apr. 2024, doi: 10.1126/scirobotics.adi8022

    doi:10.1126/scirobotics.adi8022 2024

  24. [24]

    ToddlerBot: Open-Source ML-Compatible Humanoid Platform for Loco-Manipulation

    H. Shi, W. Wang, S. Song and C. K. Liu, “ToddlerBot: Open-Source ML-Compatible Humanoid Platform for Loco-Manipulation”, 2025in arXiv, doi: 10.48550/arXiv.2502.00893

    work page doi:10.48550/arxiv.2502.00893

  25. [25]

    A Virtual Reality Framework for Safe Global Bimanual Telepresence

    N. Kosanovic and J. Vaz, “A Virtual Reality Framework for Safe Global Bimanual Telepresence ”,2025 IEEE 21st International Conference on Automation Science and Engineering (CASE), Los Angeles, CA, USA, 2025

    2025

  26. [26]

    ROS: an open-source Robot Operating System,

    M. Quigleyet al., “ROS: an open-source Robot Operating System,” 2009 ICRA Workshop on Open Source Software, 2009

    2009

  27. [27]

    Unity-Robotics-Hub,

    “Unity-Robotics-Hub,” GitHub repository, 2022. [Online]. Available: https://github.com/Unity-Technologies/Unity-Robotics-Hub

    2022

  28. [28]

    Memetic Evolution for Generic Full-Body Inverse Kinematics in Robotics and Animation,

    S. Starke, N. Hendrich and J. Zhang, “Memetic Evolution for Generic Full-Body Inverse Kinematics in Robotics and Animation,”in IEEE Transactions on Evolutionary Computation, vol. 23, no. 3, pp. 406- 420, June 2019, doi: 10.1109/TEVC.2018.2867601

    arXiv 2019

  29. [29]

    Converting a fisheye image into a panoramic, spherical or perspective projection,

    P. Bourke, “Converting a fisheye image into a panoramic, spherical or perspective projection,” 2016. Accessed: Feb. 19, 2025. [Online]. Available: https://paulbourke.net/dome/fish2/

    2016

  30. [30]

    From Universe to Metaverse: A Leap Into Virtual Collaboration at NASA JPL,

    S. -H. Berndt et al., “From Universe to Metaverse: A Leap Into Virtual Collaboration at NASA JPL,” inIEEE Transactions on Industrial Cyber-Physical Systems, vol. 1, pp. 287-306, 2023, doi: 10.1109/TICPS.2023.3327948

    arXiv 2023

  31. [31]

    Low-Latency Immersive 6D Televisualization with Spherical Rendering,

    M. Schwarz and S. Behnke, “Low-Latency Immersive 6D Televisualization with Spherical Rendering,”2020 IEEE- RAS 20th International Conference on Humanoid Robots (Humanoids), Munich, Germany, 2021, pp. 320-325, doi: 10.1109/HUMANOIDS47582.2021.9555797

    arXiv 2020

  32. [32]

    J. C. Vaz, A. Dave, N. Kassai, N. Kosanovic and P. Y . Oh, ”Immersive Auditory-Visual Real-Time Avatar System of ANA Avatar XPRIZE Finalist Avatar-Hubo,”2022 IEEE International Conference on Advanced Robotics and Its Social Impacts (ARSO), Long Beach, CA, USA, 2022, pp. 1-6, doi: 10.1109/ARSO54254.2022.9802964

    arXiv 2022

  33. [33]

    Sim-to-Real: Learning Agile Locomotion For Quadruped Robots

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, and V . Vanhoucke, “Sim-to-Real: Learning Agile Locomotion For Quadruped Robots”, 2018, arXiv:1804.10332

    Pith/arXiv arXiv 2018

  34. [34]

    Isaac Sim-to-Real: Reinforcement Learning based Locomotion for Quadrupeds,

    J. Dowdy and J. Chagas Vaz, “Isaac Sim-to-Real: Reinforcement Learning based Locomotion for Quadrupeds,” in2025 IEEE 21st International Conference on Automation Science and Engineering (CASE), Los Angeles, CA, USA, 2025, pp. 2194–2199, doi: 10.1109/CASE58245.2025.11163761

    arXiv 2025

  35. [35]

    Extended Friction Models for the Physics Simulation of Servo Actuators

    M. Duclusaud, G. Passault, V . Padois and O. Ly, “Extended Friction Models for the Physics Simulation of Servo Actuators”, 2025, arXiv:2410.08650

    arXiv 2025