Pith. sign in

REVIEW 5 major objections 5 minor 33 references

Realizing Text-Driven Motion Generation on NAO Robot: A Reinforcement Learning-Optimized Control Pipeline

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

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

Pith's one-line read The paper claims that a text prompt can drive a real NAO robot through a learned angle-mapping network and a reinforcement-learned control policy, and it demonstrates the pipeline on the physical hardware.

desk verdict A real engineering integration that moves a physical NAO from text prompts, but the 'successful transfer' claim is broader than the evidence: no baseline, no actuator calibration, and the reported failures narrow the result. read the letter →

arxiv 2506.05117 v1 pith:MB2R72JK submitted 2025-06-05 cs.RO

Zihan Xu , Mengxian Hu , Kaiyan Xiao , Qin Fang , Chengju Liu , Qijun Chen This is my paper
classification cs.RO
keywords text-drivenmotiongenerationhumanoidrobotretargetingreinforcementlearningsim-to-realtransferNAOdomainrandomizationanglesignalnetwork
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a physical humanoid robot can execute motions specified only by a natural-language text prompt, with no motion-capture or demonstration data. It does this by splitting the problem into a learned retargeting step and a learned stabilization step: a network that converts normalized limb poses into NAO joint angles, followed by a reinforcement-learned whole-body controller that tracks those angles while keeping the robot upright. The authors report the pipeline running on a real NAO robot, reproducing upper-body gestures, recovering balance after a push, and producing a partial walk. If correct, this makes text a practical command interface for humanoid motion and removes a major cost barrier to imitation learning for robots.

What carries the argument

The load-bearing device is the angle signal network trained with NPR Loss, operating on a scale-invariant limb representation. For each arm and leg, the position feature is the vector from the proximal joint (shoulder or hip) to the distal joint (wrist or ankle), divided by the limb length; the rotation feature is a quaternion derived from the vector between two distal landmarks relative to the robot's horizontal default direction. The network is an MLP with two hidden layers, followed by a sigmoid output layer that enforces joint limits. The companion mechanism is the whole-body control policy: a PPO-trained MLP that outputs joint-position targets for a PD torque law, with a reward function combining pose-tracking error, contact-force safety, motion smoothness, and flat-torso and flat-feet preferences, and with termination on tilting more than 60 degrees. A mirrored HipYaw constraint is applied because the NAO's left and right HipYaw joints are mechanically linked.

What would settle it

Measure the physical NAO's joint tracking response under the trained policy while perturbing the actuator stiffness and damping values by plus or minus ten percent from the simulator's settings. If small perturbations cause visible oscillation or tracking error, the simulator's actuator model is not faithful, and the demonstrated transfer holds only at the single hand-tuned operating point.

Watch

Extended reading notes

Core claim

The central claim is that text-driven human motion transfers to a real humanoid robot through a two-stage pipeline, and that the transfer can be made robust without per-motion human input. The first stage is an angle signal network that maps a 28-dimensional feature vector, built from scale-normalized limb displacement vectors and rotation quaternions for the four limbs, into 21 NAO joint angles; it is trained with a norm-position and rotation (NPR) loss that weights limb-length-scaled position error and quaternion-based rotation error. The second stage is a whole-body reinforcement-learning policy, trained with PPO, which consumes the joint-angle targets plus the robot's gravity estimate and current joint state, and outputs PD position commands. Training uses domain randomization of friction, mass, and pushes, while actuator gains are deliberately kept fixed because the physical robot's joint controller is highly sensitive to them. The authors state that the trained policy was deployed on the physical NAO with roughly one millisecond of inference time, and that it reproduced waving, hand-touching, violin-playing, boxing, squatting, balance recovery after a push, and a non-lifting walking attempt.

Load-bearing premise

The paper assumes that the actuator parameters tuned in the simulator (maximum torque, speed, stiffness, damping) and the domain-randomization ranges faithfully represent the physical NAO's dynamics; no quantitative hardware measurement of these values is given, and the real joint controller is described as highly sensitive to them.

Editorial extensions

If this is right

  • Text prompts become a practical command interface for humanoid motion, removing the cost of motion capture or pose estimation for each new action.
  • The scale-normalized limb representation and NPR loss could be reused to retarget motion to other humanoid platforms with different limb lengths and joint limits.
  • Because the RL controller treats joint-angle targets as references rather than hard constraints, the same pipeline can absorb pushes and maintain balance away from the reference motion.
  • The frame-wise action sampling limits the system to quasi-static motions; adding temporal coherence and velocity conditioning would be needed to reach fully dynamic locomotion.

Reading between the lines

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

  • The normalized limb features are agnostic to body proportions, so the same angle signal network could in principle be trained for heterogeneous robot models, at the cost of more diverse training data.
  • The failure of full temporal sequences to transfer suggests the simulator's actuator or ground-contact dynamics diverge from the real NAO at higher speeds; testing sequential commands with a slowed time scale could isolate the cause.
  • The deliberate exclusion of actuator-gain randomization points to a sensitivity bottleneck; a natural extension is to measure real joint step responses and match simulator gains to those measurements rather than tuning by hand.
  • The walking failure indicates that self-locomotion requires explicit foot-lift and base-velocity objectives, since pose-only rewards did not produce them.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes an end-to-end pipeline for text-driven motion generation on the physical NAO humanoid: a text-conditioned diffusion model (MLCT) generates human motion, an angle signal network trained with the proposed NPR loss maps normalized limb representations to NAO joint angles, and a PPO policy trained in IsaacLab/IsaacSim with domain randomization tracks those joint targets while preserving balance. The policy is exported to ONNX and deployed on a real NAO via ROS2. The authors report qualitative demonstrations of several upper-body and lower-body motions, with explicit acknowledgment of failures in forward walking and in fully reproducing a squat-pick-up motion.

Significance. If fully substantiated, the paper would make a useful contribution by demonstrating that text-driven human motion can be transferred to a small physical humanoid without motion capture, using a relatively accessible integration of modern generative models, RL-based control, and sim-to-real techniques. The manuscript has clear strengths: it provides an open-sourced NAO simulation model and control stack, it reports real-robot deployment rather than simulation-only results, and it candidly documents failure cases and limitations, including the failure to transfer temporally coherent motion sequences and the inability to produce foot-lifting walking. These strengths are real and should be credited. However, the central claim of 'successfully transferring text-driven human motion to a real humanoid robot NAO' is not yet quantitatively established: the evaluation is largely qualitative and self-referential, the sim-to-real parameters are unvalidated against the hardware, and there are no baselines or ablations to attribute the observed behavior to the proposed components.

major comments (5)
  1. [§V-B, Fig. 4] The quantitative evidence for imitation fidelity is self-referential. In Fig. 4, the 'reference trajectory' is the output of the paper's own angle signal network, not an external measurement of the text-described human motion, and the RL reward 'DoF target' in Table II rewards tracking of that same joint target sequence. As a result, the reported trajectory comparison does not establish that the robot reproduces the intended human motion; it only shows that the robot tracks the authors' joint commands. Please provide an external evaluation metric, such as the normalized end-effector error between the generated human skeleton motion and the recorded physical robot motion, or at least a comparison against the raw diffusion-model output before the angle signal network, reported over multiple trials with means and variances.
  2. [§IV-D, Table I] The sim-to-real transfer claim depends critically on the Table I actuator parameters (max torque, max speed, stiffness, damping), but the paper provides no calibration or validation experiment showing that these values reproduce the physical NAO's joint dynamics. The paper itself states that the NAO joint controller is 'highly sensitive to simulation parameters' and that actuator gains were deliberately not randomized, which makes validation all the more necessary. Moreover, the units and semantics of 'stiffness' are ambiguous: in the IsaacSim PD controller they may be physical gains, whereas NAO's own low-level stiffness setting is typically a dimensionless value in [0,1], and the paper does not clarify the mapping. Please either document the measurement procedure for Table I or provide a quantitative comparison of simulated versus real joint step responses or torque-position curves.
  3. [§V-B, Discussions and limitations] The abstract and conclusion claim successful transfer of text-driven human motion, but the paper's own reported results are substantially narrower. Section V-B describes the forward-walking case as a failure because the robot does not lift its feet, the squat-pick-up motion is only partially reproduced (one hand rather than both in front of the chest), and the attempt to train on temporally sequential joint commands failed to transfer to the physical robot. These acknowledged limitations are laudable in themselves, but they contradict the unqualified success claim. Please either restrict the central claim to stationary and upper-body motions, or add experimental evidence that the pipeline handles dynamic locomotion tasks successfully.
  4. [§V, general evaluation] The evaluation lacks baselines and ablations, so the contribution of the proposed components is not established. In particular, there is no comparison of the full pipeline against direct execution of the angle signal network's joint commands without the RL policy, no ablation of the NPR Loss (Eq. 8) against alternatives such as joint-angle MSE or inverse-kinematics retargeting, and no study of the sensitivity to the reward weights in Table II or the NPR loss weights in Eq. 8. Without such comparisons, the observed real-robot behavior could plausibly be produced by the angle signal network itself or by the NAO's built-in stabilization rather than by the trained RL policy. Please add these control experiments.
  5. [§V-A, Fig. 7] The simulated experiments are reported qualitatively. There are no learning curves, per-task success rates, numerical tracking errors, or quantitative sim-to-real gap statistics. Since the paper's method is built on RL training in IsaacSim, the reader needs at least a table or plot showing training convergence and the distribution of terminal conditions (falls, timeouts) across the evaluated motions. A quantitative comparison of joint-angle tracking errors in simulation versus on the physical robot would also directly support the sim-to-real transfer claim.
minor comments (5)
  1. [Eq. (5)] The word 'demotes' is a typo and should be 'denotes'.
  2. [Table I] The table heading would benefit from explicit units for stiffness and damping, and from a sentence clarifying whether these values are IsaacSim PD gains or NAO's internal stiffness/damping parameters.
  3. [Fig. 4 caption] The caption refers to a 'heatmap visualization of error distribution with mean and standard deviation', but the error metric is never defined. Please define the error and how the mean and standard deviation are computed.
  4. [§IV-D] The phrase 'The robot’s actuator are driven by four types of motors' contains a subject-verb agreement error; it should be 'The robot's actuators are driven'.
  5. [§V-A] The abbreviation MLCT is used without definition at first use; please expand it as in reference [22] when it is introduced in the experiments.

Circularity Check

1 steps flagged · score 6.0 of 10

Mapping evaluation is self-referential: the reference trajectory is the angle signal network's own output, which is also the RL reward target.

  1. fitted input called prediction [Section III-B (Fig. 4); Section IV-B / Table II (DoF target reward)]
    "Fig. 4 illustrates the reference and the measured trajectory of the robot's arm endpoint while performing a waving motion. The reference trajectory is obtained by mapping generated motion data through the angle signal network, while the measured trajectory is captured from the physical robot. Notably, the proposed NPR Loss accurately captures the spatial features of the mapped motion, enabling precise joint sequence commands that enhance the performance of reinforcement learning."

    The reference trajectory is the output of the same angle signal network being evaluated, and the RL policy is trained with a reward that directly penalizes deviation from those same q_target commands (Table II: 'DoF target', exp(-Σ|q_t - q_target|/5.0), weight 10.0; Section IV-B: 'L2-norm of the difference between the joint commands output by the angle signal net and the actual executed joint positions'). Thus the measured-vs-reference agreement in Fig. 4 reduces to the policy tracking its own training target; it does not independently test whether the mapped motion preserves the text-described human motion. The NPR Loss claim is supported by comparing the robot to the network's own output, a closed loop.

full rationale

The central derivation is mostly self-contained: the text-to-motion generator is an externally benchmarked prior model (MLCT, trained on HumanML3D), the angle signal network is supervised by NPR Loss against normalized human-motion features, and the RL policy is trained to track those joint commands. None of these steps is circular by construction. The circularity is confined to the quantitative evaluation of the mapping: Fig. 4 compares the real robot's endpoint trajectory to a reference obtained by mapping generated motion through the same angle signal network, and the RL reward (Table II 'DoF target') is exactly the tracking of that network's joint commands. Measured-vs-reference agreement therefore demonstrates that the policy tracks its own training target, not that the text-described human motion survived the mapping. This is a 'prediction reduces by construction' situation, warranting a 6 on the circularity scale. The real-robot videos (waving, boxing, playing violin) provide partially independent evidence that recognizable gestures are produced, so the circularity is partial rather than total. The self-citation of MLCT is not load-bearing because MLCT has external HumanML3D benchmarks. The admitted failure cases (walking, partial squat-pick-up) and relaxed lower-body tracking further narrow the deployment claim but are correctness risks, not additional circularity. Table I actuator parameters are unvalidated, but no circular step is exhibited for them.

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

The central claim rests on several hand-tuned parameters (reward weights, actuator settings, randomization ranges) and on domain assumptions about the sufficiency of the normalized limb representation and the fidelity of the simulation model. The use of the authors' own MLCT model as the motion generator adds a self-citation dependency, though that model is pretrained and used as an input.

free parameters (4)
  • Reward weights (Table II) = Multiple hand-set weights (e.g., -8.0e-4, -2.0e-3, -2.5e-7, -2.0e-2, -6.5e-4, -1.2, 0.3, -1.0, 10.0)
    No sensitivity analysis is provided; the success of the learned policy depends on these values.
  • NPR loss weights omega_trans and omega_quat (Eq. 8) = not reported
    The relative weighting of translational and rotational terms is chosen but never listed or ablated.
  • Joint actuator parameters (Table I) = max torque, max speed, stiffness, damping per joint type
    Fine-tuned to make the IsaacSim model match the real NAO; no validation data are given.
  • Domain randomization ranges (Table III) = friction U(0.3,1.1) / U(0.2,0.7), mass offsets, push interval 4-6 s at 0.2 m/s
    Set based on 'estimates of uncertainty' rather than measured distributions; their adequacy is critical for sim-to-real transfer.
assumptions (4)
  • domain assumption The pretrained MLCT model [22] generates motion sequences whose semantics match the input text.
    The whole pipeline starts from MLCT outputs; the paper does not independently evaluate the correctness of those motions for the prompts used.
  • domain assumption Human poses can be reduced to four limb chains with norm-position and rotation features that preserve enough information for imitation.
    Section III-B defines the mapping this way; if important motion cues (e.g., torso orientation, contact) are lost, the robot's imitation will be wrong.
  • domain assumption The IsaacSim model with Table I parameters and Table III randomization is a faithful enough proxy for the physical NAO.
    Section IV-D states actuator gains are not randomized because the real controller is sensitive, implying high fidelity is needed, but no quantitative model identification is reported.
  • standard math PPO and standard RL machinery from IsaacLab work for whole-body tracking with the provided reward.
    The paper relies on PPO [33] and the IsaacLab framework without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Realizing Text-Driven Motion Generation on NAO Robot: A Reinforcement Learning-Optimized Control Pipeline." pith.science (2026). https://pith.science/paper/MB2R72JK

@misc{pith2026250605117,
  author       = {Pith},
  title        = {Pith review of: Realizing Text-Driven Motion Generation on NAO Robot: A Reinforcement Learning-Optimized Control Pipeline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MB2R72JK}},
  note         = {Machine review of arXiv:2506.05117}
}
read the original abstract

Human motion retargeting for humanoid robots, transferring human motion data to robots for imitation, presents significant challenges but offers considerable potential for real-world applications. Traditionally, this process relies on human demonstrations captured through pose estimation or motion capture systems. In this paper, we explore a text-driven approach to mapping human motion to humanoids. To address the inherent discrepancies between the generated motion representations and the kinematic constraints of humanoid robots, we propose an angle signal network based on norm-position and rotation loss (NPR Loss). It generates joint angles, which serve as inputs to a reinforcement learning-based whole-body joint motion control policy. The policy ensures tracking of the generated motions while maintaining the robot's stability during execution. Our experimental results demonstrate the efficacy of this approach, successfully transferring text-driven human motion to a real humanoid robot NAO.

Figures

Figures reproduced from arXiv: 2506.05117 by the authors.

Figure 1
Figure 1. Visualization of NAO robot: Appearance (left) and collision model [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of mapping text-driven human motion to humanoids: (a) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Skeletal keypoints derived from human motion data and mapped [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: The architecture of the angle signal net. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Human motion animation using processed outputs of the diffusion [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Time sequence snapshots of real NAO motion control, including (a) waving right hand, (b) touching the left hand with right hand, (c) playing the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 14 canonical work pages

  1. [1]

    Real-time imitation of human whole-body motions by humanoids,

    J. Koenemann, F. Burget, and M. Bennewitz, “Real-time imitation of human whole-body motions by humanoids,” in2014 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2014, pp. 2806–2812

    work page 2014

  2. [2]

    Motionclip: Exposing human motion generation to clip space,

    G. Tevet, B. Gordon, A. Hertz, A. H. Bermano, and D. Cohen-Or, “Motionclip: Exposing human motion generation to clip space,” in European Conference on Computer Vision. Springer, 2022, pp. 358– 374

    2022

  3. [4]

    Motiondiffuse: Text-driven human motion generation with diffusion model,

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu, “Motiondiffuse: Text-driven human motion generation with diffusion model,”arXiv preprint arXiv:2208.15001, 2022

    arXiv 2022

  4. [5]

    Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation,

    J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu, “Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 3383–3393

    work page 2021

  5. [6]

    Learning human-to-humanoid real-time whole-body teleoperation,

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,” arXiv preprint arXiv:2403.04436, 2024

    arXiv 2024

  6. [7]

    Inverse kinemat- ics of a nao humanoid robot using kinect to track and imitate human motion,

    S. Mukherjee, D. Paramkusam, and S. K. Dwivedy, “Inverse kinemat- ics of a nao humanoid robot using kinect to track and imitate human motion,” in2015 International Conference on Robotics, Automation, Control and Embedded Systems (RACE). IEEE, 2015, pp. 1–7

    work page 2015

  7. [8]

    Adaptive whole-body manipulation in human-to-humanoid multi-contact motion retargeting,

    K. Otani and K. Bouyarmane, “Adaptive whole-body manipulation in human-to-humanoid multi-contact motion retargeting,” in2017 IEEE-RAS 17th International Conference on Humanoid Robotics (Humanoids). IEEE, 2017, pp. 446–453

    work page 2017

  8. [9]

    Adherent: Learning human-like trajectory generators for whole-body control of humanoid robots,

    P. M. Viceconte, R. Camoriano, G. Romualdi, D. Ferigo, S. Dafarra, S. Traversaro, G. Oriolo, L. Rosasco, and D. Pucci, “Adherent: Learning human-like trajectory generators for whole-body control of humanoid robots,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2779–2786, 2022

    work page 2022

Show all 33 references
  1. [10]

    Human motion retargeting to pepper humanoid robot from uncalibrated videos using human pose estimation,

    H. Khalil, E. Coronado, and G. Venture, “Human motion retargeting to pepper humanoid robot from uncalibrated videos using human pose estimation,” in2021 30th IEEE International Conference on Robot & Human Interactive Communication (RO-MAN). IEEE, 2021, pp. 1145–1152

    2021

  2. [11]

    Learning stable nonlinear dynamical systems with gaussian mixture models,

    S. M. Khansari-Zadeh and A. Billard, “Learning stable nonlinear dynamical systems with gaussian mixture models,”IEEE Transactions on Robotics, vol. 27, no. 5, pp. 943–957, 2011

    2011

  3. [12]

    A deep learning framework for character motion synthesis and editing,

    D. Holden, J. Saito, and T. Komura, “A deep learning framework for character motion synthesis and editing,”ACM Transactions on Graphics (TOG), vol. 35, no. 4, pp. 1–11, 2016

    2016

  4. [13]

    Action2motion: Conditioned generation of 3d human motions,

    C. Guo, X. Zuo, S. Wang, S. Zou, Q. Sun, A. Deng, M. Gong, and L. Cheng, “Action2motion: Conditioned generation of 3d human motions,” inProceedings of the 28th ACM International Conference on Multimedia, 2020, pp. 2021–2029

    2020

  5. [14]

    Action-conditioned 3d human motion synthesis with transformer vae,

    M. Petrovich, M. J. Black, and G. Varol, “Action-conditioned 3d human motion synthesis with transformer vae,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 985–10 995

    2021

  6. [15]

    Generating diverse and natural 3d human motions from text,

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng, “Generating diverse and natural 3d human motions from text,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5152–5161

    2022

  7. [16]

    Learning diverse stochastic human-action generators by learning smooth latent transitions,

    Z. Wang, P. Yu, Y . Zhao, R. Zhang, Y . Zhou, J. Yuan, and C. Chen, “Learning diverse stochastic human-action generators by learning smooth latent transitions,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 12 281–12 288

    2020

  8. [17]

    Deep video generation, prediction and completion of human action sequences,

    H. Cai, C. Bai, Y .-W. Tai, and C.-K. Tang, “Deep video generation, prediction and completion of human action sequences,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 366–382

    2018

  9. [18]

    Executing your commands via motion diffusion in latent space,

    X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu, “Executing your commands via motion diffusion in latent space,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 000–18 010

    2023

  10. [19]

    Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs,

    P. Jin, Y . Wu, Y . Fan, Z. Sun, Y . Wei, and L. Yuan, “Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs,”arXiv preprint arXiv:2311.01015, 2023

    2023 arXiv

  11. [20]

    Human motion diffusion model,

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano, “Human motion diffusion model,” inInternational Confer- ence on Learning Representations, 2023

    2023

  12. [21]

    Priority-centric human motion generation in discrete latent space,

    H. Kong, K. Gong, D. Lian, M. B. Mi, and X. Wang, “Priority-centric human motion generation in discrete latent space,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 14 806–14 816

    2023

  13. [22]

    Efficient text-driven motion generation via latent consistency training,

    M. Hu, M. Zhu, X. Zhou, Q. Yan, S. Li, C. Liu, and Q. Chen, “Efficient text-driven motion generation via latent consistency training,”arXiv preprint arXiv:2405.02791, 2024

    2024 arXiv

  14. [23]

    Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,”ACM Transactions On Graphics (TOG), vol. 37, no. 4, pp. 1–14, 2018

    2018

  15. [24]

    Sim-to-real: Learning agile locomotion for quadruped robots,

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bo- hez, and V . Vanhoucke, “Sim-to-real: Learning agile locomotion for quadruped robots,”arXiv preprint arXiv:1804.10332, 2018

    2018 arXiv

  16. [25]

    Ai-cpg: Adaptive imitated central pattern generators for bipedal locomotion learned through reinforced reflex neural networks,

    G. Li, A. Ijspeert, and M. Hayashibe, “Ai-cpg: Adaptive imitated central pattern generators for bipedal locomotion learned through reinforced reflex neural networks,”IEEE Robotics and Automation Letters, 2024

    2024

  17. [26]

    Robust feed- back motion policy design using reinforcement learning on a 3d digit bipedal robot,

    G. A. Castillo, B. Weng, W. Zhang, and A. Hereid, “Robust feed- back motion policy design using reinforcement learning on a 3d digit bipedal robot,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 5136–5143

    2021

  18. [27]

    Impedance learning-based adaptive force tracking for robot on unknown terrains,

    Y . Li, L. Zheng, Y . Wang, E. Dong, and S. Zhang, “Impedance learning-based adaptive force tracking for robot on unknown terrains,” IEEE Transactions on Robotics, 2025

    2025

  19. [28]

    Sim-to-real transfer in deep reinforcement learning for robotics: a survey,

    W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: a survey,” in2020 IEEE symposium series on computational intelligence (SSCI). IEEE, 2020, pp. 737–744

    2020

  20. [29]

    Learning agile and dynamic motor skills for legged robots,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots,”Science Robotics, vol. 4, no. 26, p. eaau5872, 2019

    2019

  21. [30]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automa- tion Letter...

    2023

  22. [31]

    Generating diverse and natural 3d human motions from text,

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng, “Generating diverse and natural 3d human motions from text,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 5152–5161

    2022

  23. [32]

    Smpl: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “Smpl: A skinned multi-person linear model,” inSeminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023, pp. 851–866

    2023

  24. [33]

    Proximal policy optimization algorithms,

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

    2017 arXiv

  25. [34]

    An efficient orientation filter for inertial and inertial/magnetic sensor arrays,

    S. Madgwicket al., “An efficient orientation filter for inertial and inertial/magnetic sensor arrays,”Report x-io and University of Bristol (UK), vol. 25, pp. 113–118, 2010

    2010

Pith tools

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