Pith. sign in

REVIEW 3 major objections 38 references

A kinematic normalizing flow plus a low-level imitation filter lets a quadruped with an arm exploit whole-body redundancy for precise loco-manipulation.

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 ↗

A hierarchical controller using a kinematic normalizing flow for partial inverse-kinematics redundancy plus low-level imitation yields 4.5 cm / 0.14 rad end-effector tracking while walking on a real quadruped-arm platform.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Clean hierarchical loco-manipulation controller that turns partial kinematic redundancy into real tracking gains on an 18-DoF platform; soft spot is the usual kinematic-to-dynamics hand-off, not a structural flaw. the 3 major comments →

arxiv 2607.11041 v1 pith:IU5VXHEP submitted 2026-07-13 cs.RO

PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinematic Embeddings

Zhengmao He , Moonkyu Jung , Hyeongjun Kim , Jiseong Lee , Hui Zhang , Jemin Hwangbo , Jie Song This is my paper
classification cs.RO
keywords loco-manipulationwhole-body controlnormalizing flowkinematic embeddinghierarchical RLquadrupedal manipulatorredundancy exploitationsim-to-real
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.

The reading

High-DoF loco-manipulation is hard because the joint search space is huge, locomotion and manipulation goals conflict, and most methods settle for one feasible solution instead of using the many redundant ones a floating-base arm system actually has. This paper claims the problem can be split cleanly: first learn a compact distribution over partial kinematic solutions (torso height/roll/pitch plus arm joints) with a normalizing flow trained on 170 million forward-kinematics samples; then let a high-level policy navigate that latent space while a low-level RL policy turns the chosen partial references into physically consistent full-body commands. On a real quadruped with a six-DoF arm the resulting system tracks end-effector poses to roughly 4.5 cm and 0.14 rad while still following base velocity commands, covering a larger usable workspace than prior whole-body baselines. A sympathetic reader cares because the split turns an intractable high-DoF RL problem into something that is both more accurate and more transferable to real mobile tasks such as cart pulling, sweeping, and plug-in.

Core claim

Decomposing whole-body loco-manipulation into partial kinematic reference generation (via a Kinematic Normalizing Flow that encodes redundant inverse-kinematic solutions) and low-level imitation control lets a high-level policy exploit system redundancy while a low-level policy enforces dynamic feasibility, yielding substantially higher end-effector and base-velocity tracking accuracy and larger feasible workspace than prior end-to-end or decoupled baselines on a quadrupedal mobile manipulator.

What carries the argument

Kinematic Normalizing Flow (KNF): a conditional normalizing flow trained on 170 million partial joint-pose pairs that maps a desired end-effector pose to a distribution of feasible partial references (torso height/roll/pitch + 6 arm DoFs); the high-level controller acts in its latent space and the low-level controller filters and executes the decoded references.

Load-bearing premise

That a purely kinematic model of partial joint configurations, filtered only by a low-level imitation policy, is enough to recover dynamically stable full-body motions for whatever latent code the high-level policy chooses.

What would settle it

On the same hardware, measure whether end-effector and base-velocity errors remain under the reported thresholds when the high-level policy is forced to sample latent codes far from the training distribution of the KNF (or when contact-rich tasks require leg configurations the partial kinematic dataset never saw).

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

Share X Bluesky LinkedIn Reddit HN

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 / 0 minor

Summary. The paper presents PAKE, a hierarchical whole-body loco-manipulation framework for a quadruped with a six-DoF arm. A Kinematic Normalizing Flow (KNF) is trained on ~170M partial forward-kinematics samples (torso height/roll/pitch + arm joints) to embed a distribution of inverse-kinematic solutions conditioned on end-effector pose. A high-level RL policy navigates the KNF latent space to select partial references that exploit redundancy; a low-level imitation RL policy tracks those partial references plus base velocity commands and produces full-DoF actuator targets via PD control. Simulation comparisons against VBC and Umi-on-Legs (plus an HLC ablation) report lower pose and velocity tracking errors and larger feasible workspace volumes. Hardware evaluation over 24 episodes of 8 tasks reports mean end-effector errors of ~4.5 cm / 0.14 rad and base velocity errors of ~0.1 m/s / 0.01 rad/s.

Significance. If the reported accuracy and workspace gains hold under broader conditions, PAKE is a useful practical contribution: it makes kinematic redundancy an explicit, navigable action space rather than leaving RL to discover a single feasible solution in a high-DoF joint space. Strengths include a large, independently generated kinematic dataset; a clean hierarchy that separates redundancy-aware planning from dynamic execution; quantitative sim ablations (Table II, (-)HLC); workspace volume metrics (Table III); and multi-task hardware numbers with motion capture (Table IV). The approach is incremental relative to prior hierarchical loco-manipulation RL, but the KNF formulation and the demonstrated combination of mobile base tracking with 6D end-effector accuracy are of clear interest to the field.

major comments (3)
  1. Abstract and §IV-D claim hardware performance “outperforming competitive baselines,” yet Table IV and the hardware section report only PAKE metrics (24 episodes / 8 tasks). No hardware numbers for VBC, Umi-on-Legs, or other cited methods appear. Either add matched hardware baselines on the same task set or restrict the outperformance claim to the simulation suite (Tables II–III), where baselines are actually evaluated.
  2. §III-A and §III-C state that KNF references “may violate physical constraints” and that the LLC “filters infeasible references,” which is load-bearing for the claim that latent navigation plus imitation substitutes for explicit dynamics/contact planning. The manuscript provides no quantitative support (e.g., fraction of HLC-selected latents that produce large tracking residuals, joint-limit or tip-over events, or recovery behavior). A short analysis or failure-case characterization would make the kinematic-to-dynamics assumption falsifiable rather than asserted.
  3. Table II / §IV-B: Umi-on-Legs is evaluated without chassis velocity tracking (marked “%” for planar and angular velocity), while PAKE and VBC are scored on simultaneous loco-manipulation. The comparison is informative for end-effector pose but is not fully fair for the paper’s central claim of coordinated locomotion + manipulation. Clarify task alignment or add a velocity-capable concurrent baseline among those already cited (e.g., Portela et al., Jung et al.).

Circularity Check

0 steps flagged

No circularity: empirical hierarchical control with KNF trained on independent FK samples and tracking evaluated against external commands/mocap.

full rationale

The paper's derivation chain is self-contained and non-circular. The KNF is fit solely to a large kinematic dataset of 170 million forward-kinematics pairs (uniform joint samples of 9 DoFs mapped to end-effector poses in the heading frame; Sec. III-B.1–2), producing a latent embedding of partial inverse-kinematic solutions. The high-level policy then navigates that latent space via RL, while the low-level imitation policy is pretrained and fine-tuned to track the resulting partial references plus velocity commands under physical simulation (Sec. III-C–D). Final claims of tracking accuracy (Tables II–IV, Figs. 4–6) and workspace volume are measured against independently sampled target trajectories and real motion-capture data, not against quantities that were fitted into the model. No equation equates a claimed prediction to a fitted constant by construction; no uniqueness theorem or ansatz is imported from overlapping-author prior work as a load-bearing premise; and the acknowledged kinematic-to-dynamics gap (KNF references may violate joint/leg/stability constraints, filtered by the LLC) is an ordinary hierarchical limitation rather than an internal reduction. The method is therefore an ordinary empirical robotics result with independent training data and external evaluation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central performance claim rests on standard RL and generative-model machinery plus a handful of hand-chosen scales and the modeling decision that 9-DoF kinematic samples plus imitation suffice for dynamic feasibility. No new physical constants or unobservable entities are postulated; the free parameters are ordinary hyper-parameters of the learning pipeline.

free parameters (4)
  • lambda_latent_scale = 0.7
    Set by hand to 0.7 to balance diversity versus quality of KNF samples (Eq. 1); directly controls the high-level action range.
  • KNF architecture (12 Glow layers, width 12, 1024-hidden coefficient nets) = 12 layers / 1024
    Chosen to model the multimodal IK distribution; reconstruction quality and latent usability depend on these sizes.
  • PD gains (kp=100, kd=1) and actor/critic sizes = kp=100, kd=1; [256x128]/[512x256x128]
    Fixed controller and network hyper-parameters that determine tracking bandwidth and policy capacity.
  • reward curriculum scales and airtime/clearance coefficients
    Multiple scalar multipliers and curriculum bounds that shape the low-level imitation objective (Table I).
axioms (4)
  • domain assumption A normalizing flow trained solely on forward-kinematics samples of 9 DoFs can represent a useful distribution over inverse-kinematic solutions for the partial system.
    Invoked in Section III-B; underpins the claim that latent navigation exploits redundancy.
  • domain assumption A low-level RL policy trained to track random torso references and velocities can filter kinematically generated but dynamically infeasible references into stable whole-body motions.
    Core of the hierarchical decomposition (Section III-C/D); if false the high-level latent actions would produce falls or large tracking errors.
  • domain assumption Standard PPO with asymmetric actor-critic and domain randomization transfers zero-shot from RaiSim to the real robot.
    Assumed throughout the hardware section; common but not guaranteed.
  • ad hoc to paper Uniform sampling of joint limits followed by parallel FK yields a representative kinematic dataset of 170 M pairs.
    Section III-B.1; density and coverage of the learned flow depend on this sampling choice.
invented entities (2)
  • Kinematic Normalizing Flow (KNF) no independent evidence
    purpose: Compact latent embedding that maps a desired end-effector pose to a distribution of partial (torso+arm) reference configurations.
    New model instance built on Glow + SoftFlow techniques; the paper treats it as the key enabler of redundancy-aware planning.
  • Partial Kinematic Embeddings (PAKE) hierarchy no independent evidence
    purpose: Overall control architecture that separates high-level latent selection from low-level dynamic execution.
    The named framework whose performance is claimed; no external validation outside this work.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinematic Embeddings." pith.science (2026). https://pith.science/paper/IU5VXHEP

@misc{pith2026260711041,
  author       = {Pith},
  title        = {Pith review of: PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinematic Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IU5VXHEP}},
  note         = {Machine review of arXiv:2607.11041}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Loco-manipulation has recently shown promising capabilities; however, achieving high-precision control, managing the high-dimensional action space induced by many degrees of freedom (DoFs), and fully exploiting the inherent redundancy of whole-body systems remain challenging. In this paper, we propose a novel whole-body control framework that effectively addresses these challenges by decomposing the complex loco-manipulation problem into partial reference motion generation and low-level imitation control. We introduce a new Kinematic Normalizing Flow (KNF) model, trained on a large-scale kinematic dataset, that generates diverse yet feasible partial reference motions. A high-level controller is then trained to navigate the KNF's latent space to exploit redundant solutions, while a low-level controller ensures physically feasible and accurate motion execution. We validate our approach on the quadrupedal robot equipped with a six-DoF robotic arm. In simulation, experimental results show that our approach significantly outperforms state-of-the-art methods in terms of tracking accuracy and feasible workspace coverage. For hardware deployment, we evaluate the system over 24 episodes across 8 different mobile loco-manipulation tasks. The system achieves end-effector pose-tracking errors of 4.5 cm and 0.14 rad, while maintaining accurate locomotion tracking with linear and angular velocity errors of 0.1 m/s and 0.01 rad/s, respectively, outperforming competitive baselines. Our method represents a practical and powerful solution for accurate and generalized whole-body loco-manipulation in high-DoF robotic systems, with promising potential for diverse downstream robotic tasks.

Figures

Figures reproduced from arXiv: 2607.11041 by Hui Zhang, HyeongJun Kim, Jemin Hwangbo, Jie Song, Jiseong Lee, Moonkyu Jung, Zhengmao He.

Figure 1
Figure 1. Figure 1: Real-world deployments of the proposed framework across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (1) We uniformly sample 6 DoF for the robotic arm and 3 DoF for the torso, collecting 170 million pairs of joint configurations [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A schematic illustration of the partial reference motion [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Violin plot distributions comparing tracking errors of differ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 3D visualization of position and orientation tracking [PITH_FULL_IMAGE:figures/full_fig_p007_5.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

38 extracted references · 1 canonical work pages

  1. [1]

    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

  2. [2]

    Learning quadrupedal locomotion on deformable terrain,

    S. Choi, G. Ji, J. Park, H. Kim, J. Mun, J. H. Lee, and J. Hwangbo, “Learning quadrupedal locomotion on deformable terrain,”Science Robotics, vol. 8, no. 74, p. eade2256, 2023

    2023

  3. [3]

    RobustDex- Grasp: Robust dexterous grasping of general objects from single-view perception,

    H. Zhang, Z. Wu, L. Huang, S. Christen, and J. Song, “RobustDex- Grasp: Robust dexterous grasping of general objects from single-view perception,”arXiv preprint arXiv:2504.05287, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    FunGrasp: Functional grasping for diverse dexterous hands,

    L. Huang, H. Zhang, Z. Wu, S. Christen, and J. Song, “FunGrasp: Functional grasping for diverse dexterous hands,” 2025

    2025

  5. [5]

    Whole- body end-effector pose tracking,

    T. Portela, A. Cramariuc, M. Mittal, and M. Hutter, “Whole- body end-effector pose tracking,” 2024. [Online]. Available: https: //arxiv.org/abs/2409.16048

    Pith/arXiv arXiv 2024

  6. [6]

    Deep whole-body control: Learning a unified policy for manipulation and locomotion,

    Z. Fu, X. Cheng, and D. Pathak, “Deep whole-body control: Learning a unified policy for manipulation and locomotion,” inConference on Robot Learning (CoRL), 2022

    2022

  7. [7]

    Visual whole-body control for legged loco-manipulation,

    M. Liu, Z. Chen, X. Cheng, Y . Ji, R. Qiu, R. Yang, and X. Wang, “Visual whole-body control for legged loco-manipulation,”The 8th Conference on Robot Learning, 2024

    2024

  8. [8]

    Learning force con- trol for legged manipulation,

    T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal, “Learning force con- trol for legged manipulation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 15 366–15 372

    2024

  9. [9]

    Learning visual quadrupedal loco-manipulation from demonstrations,

    Z. He, K. Lei, Y . Ze, K. Sreenath, Z. Li, and H. Xu, “Learning visual quadrupedal loco-manipulation from demonstrations,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 9102–9109

    2024

  10. [10]

    Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,

    P. Arm, M. Mittal, H. Kolvenbach, and M. Hutter, “Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5717–5723

    2024

  11. [11]

    Roboduet: Learning a cooperative policy for whole-body legged loco-manipulation,

    G. Pan, Q. Ben, Z. Yuan, G. Jiang, Y . Ji, S. Li, J. Pang, H. Liu, and H. Xu, “Roboduet: Learning a cooperative policy for whole-body legged loco-manipulation,”IEEE Robotics and Automation Letters, vol. 10, no. 5, pp. 4564–4571, 2025

    2025

  12. [12]

    Learning whole-body loco-manipulation for omni-directional task space pose tracking with a wheeled-quadrupedal-manipulator,

    K. Jiang, Z. Fu, J. Guo, W. Zhang, and H. Chen, “Learning whole-body loco-manipulation for omni-directional task space pose tracking with a wheeled-quadrupedal-manipulator,”IEEE Robotics and Automation Letters, vol. 10, no. 2, pp. 1481–1488, 2025

    2025

  13. [13]

    Perception-less terrain adaptation through whole body control and hierarchical optimization,

    C. Dario Bellicoso, C. Gehring, J. Hwangbo, P. Fankhauser, and M. Hutter, “Perception-less terrain adaptation through whole body control and hierarchical optimization,” in2016 IEEE-RAS 16th In- ternational Conference on Humanoid Robots (Humanoids), 2016, pp. 558–564

    2016

  14. [14]

    Dynamic locomotion in the mit cheetah 3 through convex model-predictive control,

    J. Di Carlo, P. M. Wensing, B. Katz, G. Bledt, and S. Kim, “Dynamic locomotion in the mit cheetah 3 through convex model-predictive control,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 1–9

    2018

  15. [15]

    Highly dynamic quadruped locomotion via whole-body impulse control and model predictive control,

    D. Kim, J. Di Carlo, B. Katz, G. Bledt, and S. Kim, “Highly dynamic quadruped locomotion via whole-body impulse control and model predictive control,”arXiv preprint arXiv:1909.06586, 2019

    Pith/arXiv arXiv 1909

  16. [16]

    Perceptive locomotion through nonlinear model-predictive control,

    R. Grandia, F. Jenelten, S. Yang, F. Farshidian, and M. Hutter, “Perceptive locomotion through nonlinear model-predictive control,” IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3402–3421, 2023

    2023

  17. [17]

    Per-contact iteration method for solving contact dynamics,

    J. Hwangbo, J. Lee, and M. Hutter, “Per-contact iteration method for solving contact dynamics,”IEEE Robotics and Automation Letters, vol. 3, no. 2, pp. 895–902, 2018. [Online]. Available: www.raisim.com

    2018

  18. [18]

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

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Mack- lin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac gym: High performance GPU based physics simulation for robot learning,” inThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

    2021

  19. [19]

    Learning arm-assisted fall damage reduction and recovery for legged mobile manipulators,

    Y . Ma, F. Farshidian, and M. Hutter, “Learning arm-assisted fall damage reduction and recovery for legged mobile manipulators,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 12 149–12 155

    2023

  20. [20]

    Multi- expert learning of adaptive legged locomotion,

    C. Yang, K. Yuan, Q. Zhu, W. Yu, and Z. Li, “Multi- expert learning of adaptive legged locomotion,”Science Robotics, vol. 5, no. 49, p. eabb2174, 2020. [Online]. Available: https: //www.science.org/doi/abs/10.1126/scirobotics.abb2174

    work page doi:10.1126/scirobotics.abb2174 2020

  21. [21]

    Walk these ways: Tuning robot control for generalization with multiplicity of behavior,

    G. B. Margolis and P. Agrawal, “Walk these ways: Tuning robot control for generalization with multiplicity of behavior,” in6th Annual Conference on Robot Learning, 2022. [Online]. Available: https://openreview.net/forum?id=52c5e73SlS2

    2022

  22. [22]

    Anymal parkour: Learning agile navigation for quadrupedal robots,

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter, “Anymal parkour: Learning agile navigation for quadrupedal robots,”Science Robotics, vol. 9, no. 88, p. eadi7566, 2024. [Online]. Available: https://www.science.org/doi/abs/10.1126/scirobotics.adi7566

    doi:10.1126/scirobotics.adi7566 2024

  23. [23]

    Extreme parkour with legged robots,

    X. Cheng, K. Shi, A. Agarwal, and D. Pathak, “Extreme parkour with legged robots,” inTowards Generalist Robots: Learning Paradigms for Scalable Skill Acquisition @ CoRL2023, 2023

    2023

  24. [24]

    Robot parkour learning,

    Z. Zhuang, Z. Fu, J. Wang, C. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao, “Robot parkour learning,” inConference on Robot Learning (CoRL), 2023

    2023

  25. [25]

    A unified mpc framework for whole-body dynamic locomotion and manipula- tion,

    J.-P. Sleiman, F. Farshidian, M. V . Minniti, and M. Hutter, “A unified mpc framework for whole-body dynamic locomotion and manipula- tion,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4688– 4695, 2021

    2021

  26. [26]

    Alma - articulated locomotion and manipulation for a torque-controllable robot,

    C. D. Bellicoso, K. Krämer, M. Stäuble, D. Sako, F. Jenelten, M. Bjelonic, and M. Hutter, “Alma - articulated locomotion and manipulation for a torque-controllable robot,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 8477– 8483

    2019

  27. [27]

    Roloma: Robust loco-manipulation for quadruped robots with arms,

    H. Ferrolho, V . Ivan, W. Merkt, I. Havoutis, and S. Vijayakumar, “Roloma: Robust loco-manipulation for quadruped robots with arms,” Autonomous Robots, vol. 47, no. 8, pp. 1463–1481, 2023

    2023

  28. [28]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,”Science robotics, vol. 5, no. 47, p. eabc5986, 2020

    2020

  29. [29]

    Combining learning-based locomotion policy with model-based manipulation for legged mobile manipulators,

    Y . Ma, F. Farshidian, T. Miki, J. Lee, and M. Hutter, “Combining learning-based locomotion policy with model-based manipulation for legged mobile manipulators,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2377–2384, 2022

    2022

  30. [30]

    UMI-on-legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,

    H. Ha, Y . Gao, Z. Fu, J. Tan, and S. Song, “UMI-on-legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,” in8th Annual Conference on Robot Learning, 2024. [Online]. Available: https://openreview.net/forum?id=3i7j8ZPnbm

    2024

  31. [31]

    Learning dynamic pick-and-place for a legged manipulator,

    M. Jung, J. Lee, Z. He, D. Youm, J. Mun, H. Kim, H. Oh, D. Choi, J. Hur, J. Song, and J. Hwangbo, “Learning dynamic pick-and-place for a legged manipulator,”IEEE Robotics and Automation Letters, vol. 11, no. 6, pp. 7652–7659, 2026

    2026

  32. [32]

    Ikflow: Generating diverse inverse kinematics solutions,

    B. Ames, J. Morgan, and G. Konidaris, “Ikflow: Generating diverse inverse kinematics solutions,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7177–7184, 2022

    2022

  33. [33]

    Softflow: Probabilistic framework for normalizing flow on manifolds,

    H. Kim, H. Lee, W. H. Kang, J. Y . Lee, and N. S. Kim, “Softflow: Probabilistic framework for normalizing flow on manifolds,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, pp. 16 388–16 397. [Online]. Available: https://proceedings.neurips.cc...

    2020

  34. [34]

    Glow: Generative flow with invertible 1x1 convolutions,

    D. P. Kingma and P. Dhariwal, “Glow: Generative flow with invertible 1x1 convolutions,” inAdvances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018. [Online]. Available: https://proceedings.neurips.cc/paper_ files/paper/2018/file/d139db...

    2018

  35. [35]

    Learning coordinated terrain-adaptive locomotion by imitating a cen- troidal dynamics planner,

    P. Brakel, S. Bohez, L. Hasenclever, N. Heess, and K. Bousmalis, “Learning coordinated terrain-adaptive locomotion by imitating a cen- troidal dynamics planner,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 10 335–10 342

    2022

  36. [36]

    Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion,

    G. Ji, J. Mun, H. Kim, and J. Hwangbo, “Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 4630–4637, 2022

    2022

  37. [37]

    Hierarchical reinforcement learning for precise soccer shooting skills using a quadrupedal robot,

    Y . Ji, Z. Li, Y . Sun, X. B. Peng, S. Levine, G. Berseth, and K. Sreenath, “Hierarchical reinforcement learning for precise soccer shooting skills using a quadrupedal robot,” in2022 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), 2022, pp. 1479–1486

    2022

  38. [38]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” in Advances in Neural Information Processing S...

    2019

This paper was first reviewed by grok-4.5 on July 14, 2026.