Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Learning Motion Skills with Adaptive Assistive Curriculum Force in Humanoid Robots

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A second, fading helper force trains humanoid skills 30% faster

desk verdict A2CF is a genuinely new assistive-force curriculum idea with promising directional results, but the paper's headline speedup and support-free claims exceed what the single-seed, no-error-bar evaluation actually shows. read the letter →

arxiv 2506.23125 v1 pith:SO4E7QCF submitted 2025-06-29 cs.RO

Zhanxiang Cao , Yang Zhang , Buqing Nie , Huangxuan Lin , Haoyang Li , Yue Gao This is my paper
classification cs.RO
keywords humanoidrobotsreinforcementlearningcurriculumassistiveforcemotionskillssim-to-realtransferbipedalwalkingbackflip
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 proposes A2CF, a reinforcement-learning method in which a second, assistive-force agent pushes the humanoid's pelvis during training and gradually stops as the main motion policy improves. The central claim is that this dual-agent setup, with a shrinking force bound, task-specific initial force limits, privileged information, and intermittent random masking, learns walking, dancing, and backflip policies about 30% faster than the DreamWaQ baseline, with failure rates reduced by over 40%. The final policies operate with zero assistive force at deployment, and the walking and dancing policies transfer to a physical humanoid robot without fine-tuning; the backflip is validated in a second simulator. If correct, A2CF is a reusable training-time accelerator for high-dimensional humanoid motion skills.

What carries the argument

The load-bearing object is the 6D assistive-force hypercube $B_k = \{F \in \mathbb{R}^6 \mid -\eta_{k,i} \leq F_i \leq \eta_{k,i}\}$ with half-widths $\eta_k$ that Algorithm 1 updates: the bound shrinks when $\|F_k\| < (1-\epsilon)\|\eta_k\|$ and when the skill flag says the motion is learned. A random mask $M_t$ drops the assistive force with probability $\zeta = 0.2$, preventing over-reliance, while a variational autoencoder compresses historical observations into a shared latent state used by both the motion policy and the assistive-force agent. This machinery turns a fixed external push into a state-dependent, time-varying curriculum that disappears before deployment.

What would settle it

Train the same three tasks with the masking probability $\zeta$ varied from 0 to 0.5 and measure deployment success with the assistive force fully disabled; if success collapses at $\zeta = 0.2$, or if $\zeta = 0$ matches A2CF's gains, the masking-based support-free claim is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that a motion policy trained while a jointly-optimized assistive-force agent applies a 6D force to the robot's pelvis, under a hypercube bound that shrinks automatically, converges faster and produces more reliable behavior than training without such assistance. The assistive force is a curriculum device: it starts strong, decays when the normalized applied force is small or when the skill-acquisition flag is set, and is randomly masked 20% of the time so the motion policy must sometimes act without help. Privileged information and task-specific initial force bounds shape the assistance. The consequence is that assistance can be used purely during training, yielding 'support-free' policies that need no external force at deployment.

Load-bearing premise

The paper assumes that showing the robot zero assistive force only 20% of the time, through random masking, is enough for the final policy to remain competent when assistive force is always absent at deployment.

Editorial extensions

If this is right

  • A2CF can be applied to any high-dimensional locomotion task as a training-time accelerator, since the assistive force exists only in simulation and is removed at test time.
  • The hypercube curriculum removes the need to hand-schedule assistance: the force bound decays automatically from the applied-force magnitude and a skill-acquisition indicator.
  • Random masking may generalize to other auxiliary training signals beyond physical forces, providing a general way to prevent policies from over-relying on extra guidance.
  • The real-robot walking and dancing transfers suggest the method is compatible with standard sim-to-real pipelines, producing policies that run on proprioceptive sensors alone.
  • If the reported 30% convergence speedup and over-40% failure reduction replicate across different robots, A2CF offers a practical speedup for acquiring complex motion skills.

Reading between the lines

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

  • The masking probability $\zeta$ is a key free parameter; one could expect an optimal value between full masking (no assistance) and no masking (over-reliance), so tuning it per task may improve results further.
  • The same assistive-force curriculum could be applied to manipulation skills, where a virtual force on the end effector guides early learning and fades as dexterity improves; the paper does not test this.
  • Because the backflip was validated only in a second simulator, the claim that A2CF accelerates acrobatic skills on real hardware rests entirely on the walking and dancing transfers.
  • The assistive-force reward is introduced only once the motion agent reaches 80% of the skill target; the exact placement of this threshold likely interacts strongly with the force-bound decay and masking probability, so reproducing the results may require matching this schedule.
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

3 major / 5 minor

Summary. The paper proposes A2CF, a dual-agent reinforcement learning framework in which an assistive force agent applies bounded 6-D forces to the robot's pelvis during training, with an adaptive curriculum that shrinks the force bound based on applied force magnitude and skill-completion flags, along with privileged information, task-specific initial force bounds, and random masking. The approach is evaluated on walking, backflip, and dancing for the Unitree G1 humanoid in Isaac Lab, with real-world walking and dancing demos and a Sim2Sim backflip validation in Genesis. The central claims are faster convergence (30%), lower failure rates (40%), and final policies that are robust and support-free at deployment.

Significance. If the claims hold, A2CF offers a reusable training-time accelerator for high-dimensional humanoid skill acquisition, with a principled mechanism for fading assistance. Strengths include the multi-task evaluation, ablations of PI, ID, and RM, the use of a standard simulation platform, and the attempt at real-world transfer. The main limitations are that the headline quantitative claims are not tied to any reported statistics and the support-free property is not directly tested; both should be fixable with additional experiments.

major comments (3)
  1. [Abstract; Section IV-C, Figs. 4-6] The abstract's quantitative claims of 'convergence 30% faster' and 'lowers failure rates by over 40%' are not derivable from any reported table or multi-seed statistic. In Fig. 4, A2CF reaches the maximum terrain level at approximately 4k iterations versus approximately 10k for Baseline, which is about 60% faster, not 30%. Fig. 5 shows success-rate curves but no numeric failure rates or error bars. The authors should provide a table with mean and standard deviation across multiple seeds for all tasks and baselines, and reconcile the reported percentages with these numbers.
  2. [Section III-B.5; Section IV-F; Figs. 4-6] The central claim of 'robust, support-free policies' is never directly tested. All simulation metrics are training curves in which the assistive force agent remains active throughout; the force bound decays to approximately 0.1 N, but the assistive agent is not removed. The only zero-force exposure is the random mask with probability zeta = 0.2, meaning the policy sees a completely unsupported step only 20% of the time. No post-training simulation evaluation is reported with the assistive force set exactly to zero for all steps, and the real-world results are qualitative video demonstrations with no trial counts or failure rates. The backflip was not tested on hardware at all. To support the abstract's claim, the authors should report a quantitative evaluation with the assistive force fully disabled at deployment, including real-world trial statistics.
  3. [Algorithm 1; Section III-D, 'Force Less Assi. Force' rewards] The decay of the assistive force to zero is partly engineered by the curriculum and the reward function: Algorithm 1 shrinks the bound whenever the normalized force magnitude is low or the skill flag is set, and the reward term exp[-2||F||/||eta||] penalizes large forces. The paper should distinguish between the force bound (an action-space limit) and the actual applied force, and report the actual applied force magnitude during training and at evaluation. Without this, the claim that the assistive force is 'gradually optimized to zero' (Fig. 1 caption) conflates the bound with the learned behavior and weakens the adaptivity claim.
minor comments (5)
  1. [Section IV-E] The dancing results text references 'Figure 5' but the corresponding training curves are shown in Figure 6; please correct the cross-reference.
  2. [Section IV-C] The sentence 'Since the terrain level randomly shifts to any difficulty once the robot reaches its maximum level, the average maximum terrain level is approximately 6' is confusing; please clarify the terrain curriculum and how the skill acquisition flag is determined.
  3. [Table IV] The units for the force and moment bounds are not specified; please state N and N-m explicitly.
  4. [Section III-B.5] The random mask is described as a scalar M_t in {0,1} but applied as M_t ⊙ F_t, implying a per-component mask; please clarify whether the mask is scalar or per-dimension.
  5. [Section IV-A] No hyperparameters are given for the PPO training (learning rate, clipping range, GAE lambda, etc.); please include a table for reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: A2CF's speedup and success claims are measured against baselines; the zero-force behavior is explicitly engineered, and the only self-citation is non-load-bearing.

full rationale

This is an empirical RL systems paper rather than a derivation, and its central claims are benchmark comparisons: A2CF reaches the maximum terrain level around 4k iterations versus about 10k for Baseline (Sec. IV-C), the backflip success rate exceeds 90% (Sec. IV-D), and dancing improves joint tracking (Sec. IV-E). These results are compared against DreamWaQ with matching reward terms (except assistive-force terms), curriculum, domain randomization, seeds, and network architecture (Sec. IV-B), so the speedup claim is not fitted to the reported outcome by construction. The force-to-zero behavior is not a hidden prediction: Algorithm 1 explicitly multiplies the hypercube bound by (1-delta) whenever isSkillLearned is set (lines 9-11), and the 'Force Less' reward penalizes large assistive forces, so the decay to ~0.1 N is an engineered curriculum described as such in Sec. III-B.2. No equation in the paper reduces a reported result to a parameter fitted to that same result. The only self-citation is [30], a same-group VAE-based history encoder used alongside the external DreamWaQ architecture [29]; it is an implementation component and is not load-bearing for the speed/success claims. There is a genuine evidence gap: 'support-free' robustness is not demonstrated in simulation with the assistive force removed, the real-world walking and dancing results are qualitative video demonstrations, and the backflip was not tested on hardware (Sec. IV-F). This weakens the abstract's strength but is an evidence/correctness concern, not circularity. Score 2 reflects only the minor non-load-bearing self-citation, not a circular derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several hand-tuned parameters (task-specific force bounds, curriculum rates, mask probability, skill thresholds) and on domain assumptions about simulator fidelity and the coaching analogy. The assistive force agent is an internal algorithmic entity without independent falsifiable evidence.

free parameters (6)
  • Initial force bound bounds eta_0 for walking = [40,40,10] N, [40,40,40] Nm
    Table IV: hand-designed task-specific priors; the walking task assumes vertical force demand is minimal. Ablation A2CF w/o ID shows the bounds matter.
  • Initial force bounds for backflip phases = per-phase values in Table IV, e.g. z-force 100 during Jump/Air
    Phase-dependent bounds are manually tuned; ablation A2CF w/o ID shows uniform bounds hurt performance, confirming these parameters are load-bearing.
  • Initial force bounds for dancing = [40,40,40] N and [40,40,40] Nm
    Table IV; no task-specific tuning, unlike backflip.
  • Curriculum hyperparameters epsilon, delta = epsilon=0.5, delta=0.2
    Algorithm 1; chosen by hand, they control how quickly the assistive bound shrinks or grows; no sensitivity analysis is provided.
  • Random mask probability zeta = 0.2
    Section III-B.5; controls the frequency of support-free experience and is central to the independence claim.
  • Skill acquisition thresholds = walking: max terrain level; backflip: reach land phase and remain stable; dancing: joint tracking error <0.16; force…
    These determine when the force bound is forced to decay; if mis-set, force removal could be premature.
assumptions (4)
  • domain assumption Isaac Lab and Genesis simulators faithfully model Unitree G1 dynamics to the extent needed for zero-shot sim-to-real transfer.
    The paper transfers walking and dancing policies to a real robot with no fine-tuning (Section IV-F); if simulator dynamics are not faithful, real-world results would not follow.
  • domain assumption A force applied to the pelvis is a benign teaching signal that shapes exploration without causing the motion policy to overfit to force.
    The entire framework relies on this coaching analogy (Introduction, Section III-B); the paper provides no theoretical guarantee.
  • domain assumption Jointly training the assistive force agent with PPO on the same reward does not create conflicting gradients that destabilize the motion policy.
    The method uses joint action learners and PPO (Section III-B, III-C); stability is asserted empirically, not proven.
  • domain assumption PPO with the provided hyperparameters converges in practice.
    The paper uses PPO without convergence guarantees; all claims depend on the training runs reaching the reported performance.
invented entities (1)
  • Assistive force agent
    purpose: A second policy trained jointly with the motion policy; outputs 6D spatial forces and torques applied to the robot's pelvis during simulation, with actions bounded by an adaptive hypercube.
    The agent exists only inside the training loop; its utility is demonstrated through the final task performance of the main policy, not through any independent prediction or measurement. It is a useful algorithmic device, not a physical discovery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Motion Skills with Adaptive Assistive Curriculum Force in Humanoid Robots." pith.science (2026). https://pith.science/paper/SO4E7QCF

@misc{pith2026250623125,
  author       = {Pith},
  title        = {Pith review of: Learning Motion Skills with Adaptive Assistive Curriculum Force in Humanoid Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SO4E7QCF}},
  note         = {Machine review of arXiv:2506.23125}
}
read the original abstract

Learning policies for complex humanoid tasks remains both challenging and compelling. Inspired by how infants and athletes rely on external support--such as parental walkers or coach-applied guidance--to acquire skills like walking, dancing, and performing acrobatic flips, we propose A2CF: Adaptive Assistive Curriculum Force for humanoid motion learning. A2CF trains a dual-agent system, in which a dedicated assistive force agent applies state-dependent forces to guide the robot through difficult initial motions and gradually reduces assistance as the robot's proficiency improves. Across three benchmarks--bipedal walking, choreographed dancing, and backflip--A2CF achieves convergence 30% faster than baseline methods, lowers failure rates by over 40%, and ultimately produces robust, support-free policies. Real-world experiments further demonstrate that adaptively applied assistive forces significantly accelerate the acquisition of complex skills in high-dimensional robotic control.

Figures

Figures reproduced from arXiv: 2506.23125 by the authors.

Figure 1
Figure 1. Overall Algorithm Framework. The figure illustrates the pipeline of the algorithm. In the simulation, the assistive force agent applies assistive forces to accelerate learning. With the help of the force bound curriculum, the assistive force is gradually optimized to zero. In the real￾world deployment, no assistive force is required. leads to inefficient exploration, where the robot may fail to discover effective mo… view at source ↗
Figure 2
Figure 2. FSM for Backflip. The figure shows the five states of the backflip task, with arrows indicating state transitions. The conditions for these transitions are shown along the horizontal lines. TABLE II REWARD TERMS FOR BACKFLIP TASK. Type Name Equation Weight Task Alive (Phase 0) 𝟙phase0 2.0 Down. Vel. (Phase 1) 𝟙phase1 · 𝟙contact · −𝑣𝑧 2.0 Up. Vel. (Phase 2) 𝟙phase2 · 𝑣𝑧 2.0 Ang. Vel. (Phase 2) 𝟙phase2 · −𝑤𝑦 0.5 Ang. … view at source ↗
Figure 4
Figure 4. Training Curves for the Walking Task. The figure shows the terrain level, force limit curriculum, total rewards, and raw velocity tracking reward (ranging from 0 to 1). Vertical dashed lines indicate the point at which the assistive force becomes negligible. In this task, the A2CF algorithm is compared to the Baseline and A2CF without Privileged Information (w/o PI) during training. The results, shown in [PITH_FULL… view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Simulation of Walking, backflip, and Dancing Tasks. The figure illustrates the execution of three tasks—walking, backflip, and dancing—in the simulation. The walking and backflip tasks have a 0.1 s interval between frames, while the dancing task has a 1 s interval. The…
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Training Curves for the Dancing Task. The figure shows the raw tracking rewards for base position, base orientation, and joint positions, without applying reward weights. Vertical dashed lines indicate the point at which the assistive force becomes negligible. E. Danci…
Figure 7
Figure 7. Figure 7: Sim2Sim2Real Transfer Results for Walking and Sim2Sim Results for Backflip. 1) The walking task shows Sim2Sim2Real transfer from IsaacLab (training domain) to Genesis (validation domain) and to the real-world environment (deployment domain). 2) The backflip task illust…
Figure 8
Figure 8. Figure 8: Real-World Performance of the Dancing Task. The figure shows the real-world deployment of the dancing policy, demonstrating the robot performing the dancing routine in a physical environment. where the robot performs the dance routine with high accu￾racy, demonstrating…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. GS-Playground: A High-Throughput Photorealistic Simulator for Vision-Informed Robot Learning

    cs.RO 2026-04 unverdicted novelty 6.0 of 10

    GS-Playground delivers a high-throughput photorealistic simulator for vision-informed robot learning via parallel physics integrated with batch 3D Gaussian Splatting at 10^4 FPS and an automated Real2Sim workflow for ...

  2. Athena-WBC: Capability-Aligned Policy Experts for Long-Tail Humanoid Whole-Body Control

    cs.RO 2026-07 conditional novelty 5.5 of 10

    Capability-aligned dynamic and balance experts recover residual humanoid whole-body tracking failures better than data reallocation alone, then distill into one stronger deployable controller.

  3. WARL: Wrench-Augmented Reinforcement Learning for Task-Agnostic Learning in Legged Robots

    cs.RO 2026-07 conditional novelty 4.0 of 10

    Adding a simulated torso wrench during early RL training and gradually removing it lets a quadruped learn six locomotion tasks with a shared reward, yielding a joint-only policy in simulation.

Reference graph

Works this paper leans on

39 extracted references · 14 canonical work pages · cited by 3 Pith papers

  1. [1]

    Advancements in humanoid robots: A comprehensive review and future prospects,

    Y . Tong, H. Liu, and Z. Zhang, “Advancements in humanoid robots: A comprehensive review and future prospects,” IEEE/CAA Journal of Automatica Sinica, vol. 11, no. 2, pp. 301–328, 2024

    work page 2024

  2. [2]

    Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,

    Z. Gu, J. Li, W. Shen, W. Yu, Z. Xie, S. McCrory, X. Cheng, A. Shamsah, R. Griffin, C. K. Liu et al., “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,” arXiv preprint arXiv:2501.02116 , 2025

    arXiv 2025

  3. [3]

    Real-world humanoid locomotion with reinforcement learning,

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath, “Real-world humanoid locomotion with reinforcement learning,” Science Robotics, vol. 9, no. 89, p. eadi9579, 2024

    2024

  4. [4]

    A unified and general humanoid whole-body controller for fine-grained locomotion,

    Y . Xue, W. Dong, M. Liu, W. Zhang, and J. Pang, “A unified and general humanoid whole-body controller for fine-grained locomotion,” arXiv preprint arXiv:2502.03206 , 2025

    arXiv 2025

  5. [5]

    Beamdojo: Learning agile humanoid locomotion on sparse footholds,

    H. Wang, Z. Wang, J. Ren, Q. Ben, T. Huang, W. Zhang, and J. Pang, “Beamdojo: Learning agile humanoid locomotion on sparse footholds,” arXiv preprint arXiv:2502.10363 , 2025

    arXiv 2025

  6. [6]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan et al., “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,” arXiv preprint arXiv:2502.01143, 2025

    arXiv 2025

  7. [7]

    Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi, “Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,” arXiv preprint arXiv:2406.08858, 2024

    arXiv 2024

  8. [8]

    Exbody2: Advanced expressive humanoid whole-body control,

    M. Ji, X. Peng, F. Liu, J. Li, G. Yang, X. Cheng, and X. Wang, “Exbody2: Advanced expressive humanoid whole-body control,”arXiv preprint arXiv:2412.13196, 2024

    arXiv 2024

Show all 39 references
  1. [9]

    Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,

    F. Liu, Z. Gu, Y . Cai, Z. Zhou, S. Zhao, H. Jung, S. Ha, Y . Chen, D. Xu, and Y . Zhao, “Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,” arXiv preprint arXiv:2409.20514, 2024

    2024

  2. [10]

    Balancing collective exploration and exploitation in multi-agent and multi-robot systems: A review,

    H. L. Kwa, J. Leong Kit, and R. Bouffanais, “Balancing collective exploration and exploitation in multi-agent and multi-robot systems: A review,” Frontiers in Robotics and AI , vol. 8, p. 771520, 2022

    2022

  3. [11]

    Exploration in deep reinforcement learning: A survey,

    P. Ladosz, L. Weng, M. Kim, and H. Oh, “Exploration in deep reinforcement learning: A survey,” Information Fusion , vol. 85, pp. 1–22, 2022

    2022

  4. [12]

    Wulf, Attention and motor skill learning

    G. Wulf, Attention and motor skill learning . Human Kinetics, 2007

    2007

  5. [13]

    The control of posture in newly standing infants is task dependent,

    L. J. Claxton, D. K. Melzer, J. H. Ryu, and J. M. Haddad, “The control of posture in newly standing infants is task dependent,” Journal of experimental child psychology , vol. 113, no. 1, pp. 159–165, 2012

    2012

  6. [14]

    Eye–hand coordination in the newborn

    C. V on Hofsten, “Eye–hand coordination in the newborn.” Develop- mental psychology, vol. 18, no. 3, p. 450, 1982

    1982

  7. [15]

    P. H. Werner, L. H. Williams, and T. J. Hall, Teaching children gymnastics. Human Kinetics, 2012

    2012

  8. [16]

    Characteristics of effective gymnastics coaching,

    T. Dowdell, “Characteristics of effective gymnastics coaching,” Sci- ence of gymnastics Journal , vol. 2, no. 1, pp. 15–24, 2010

    2010

  9. [17]

    Hilo: Learning whole- body human-like locomotion with motion tracking controller,

    Q. Zhang, C. Weng, G. Li, F. He, and Y . Cai, “Hilo: Learning whole- body human-like locomotion with motion tracking controller,” arXiv preprint arXiv:2502.03122, 2025

    2025 arXiv

  10. [18]

    Ad- vancing humanoid locomotion: Mastering challenging terrains with denoising world model learning,

    X. Gu, Y .-J. Wang, X. Zhu, C. Shi, Y . Guo, Y . Liu, and J. Chen, “Ad- vancing humanoid locomotion: Mastering challenging terrains with denoising world model learning,” arXiv preprint arXiv:2408.14472 , 2024

    2024 arXiv

  11. [19]

    Learning humanoid standing-up control across diverse postures,

    T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, and J. Pang, “Learning humanoid standing-up control across diverse postures,” arXiv preprint arXiv:2502.08378 , 2025

    2025 arXiv

  12. [20]

    Learning getting-up policies for real-world humanoid robots,

    X. He, R. Dong, Z. Chen, and S. Gupta, “Learning getting-up policies for real-world humanoid robots,” arXiv preprint arXiv:2502.12152 , 2025

    2025 arXiv

  13. [21]

    Whole-body humanoid robot locomotion with human reference,

    Q. Zhang, P. Cui, D. Yan, J. Sun, Y . Duan, G. Han, W. Zhao, W. Zhang, Y . Guo, A. Zhang et al. , “Whole-body humanoid robot locomotion with human reference,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 11 225– 11 231

    2024

  14. [22]

    Amp: Adversarial motion priors for stylized physics-based character con- trol,

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “Amp: Adversarial motion priors for stylized physics-based character con- trol,” ACM Transactions on Graphics (ToG), vol. 40, no. 4, pp. 1–20, 2021

    2021

  15. [23]

    Deep imitation learning for humanoid loco-manipulation through human teleoperation,

    M. Seo, S. Han, K. Sim, S. H. Bang et al., “Deep imitation learning for humanoid loco-manipulation through human teleoperation,” arXiv preprint arXiv:2309.01952, 2023

    2023 arXiv

  16. [24]

    Development of a whole-body work imitation learn- ing system by a biped and bi-armed humanoid,

    Y . Matsuura, K. Kawaharazuka, N. Hiraoka, K. Kojima, K. Okada, and M. Inaba, “Development of a whole-body work imitation learn- ing system by a biped and bi-armed humanoid,” arXiv preprint arXiv:2309.15756, 2023

    2023 arXiv

  17. [25]

    Multi-agent reinforcement learning: A review of challenges and applications,

    L. Canese, G. C. Cardarilli, L. Di Nunzio, R. Fazzolari, D. Giardino, M. Re, and S. Span `o, “Multi-agent reinforcement learning: A review of challenges and applications,” Applied Sciences, vol. 11, no. 11, p. 4948, 2021

    2021

  18. [26]

    A review of cooperation in multi-agent learning,

    Y . Du, J. Z. Leibo, U. Islam, R. Willis, and P. Sunehag, “A review of cooperation in multi-agent learning,”arXiv preprint arXiv:2312.05162, 2023

    2023 arXiv

  19. [27]

    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

  20. [28]

    Asymmetric actor critic for image-based robot learning,

    L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asymmetric actor critic for image-based robot learning,” arXiv preprint arXiv:1710.06542, 2017

    2017 arXiv

  21. [29]

    Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning,

    I. M. A. Nahrendra, B. Yu, and H. Myung, “Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 5078–5084

    2023

  22. [30]

    Robust locomotion policy with adaptive lipschitz constraint for legged robots,

    Y . Zhang, B. Nie, and Y . Gao, “Robust locomotion policy with adaptive lipschitz constraint for legged robots,” IEEE Robotics and Automation Letters, 2024

    2024

  23. [31]

    beta-vae: Learning basic visual concepts with a constrained variational framework

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework.” ICLR (Poster), vol. 3, 2017

    2017

  24. [32]

    Understanding disentangling in 𝛽-vae,

    C. P. Burgess, I. Higgins, A. Pal, L. Matthey, N. Watters, G. Des- jardins, and A. Lerchner, “Understanding disentangling in 𝛽-vae,” arXiv preprint arXiv:1804.03599 , 2018

    2018 arXiv

  25. [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

  26. [34]

    Minimizing energy consumption leads to the emergence of gaits in legged robots,

    Z. Fu, A. Kumar, J. Malik, and D. Pathak, “Minimizing energy consumption leads to the emergence of gaits in legged robots,” arXiv preprint arXiv:2111.01674, 2021

    2021 arXiv

  27. [35]

    Stage-wise reward shaping for acrobatic robots: A constrained multi-objective reinforce- ment learning approach,

    D. Kim, H. Kwon, J. Kim, G. Lee, and S. Oh, “Stage-wise reward shaping for acrobatic robots: A constrained multi-objective reinforce- ment learning approach,” arXiv preprint arXiv:2409.15755 , 2024

    2024 arXiv

  28. [36]

    Robust mo- tion in-betweening,

    F. G. Harvey, M. Yurick, D. Nowrouzezahrai, and C. Pal, “Robust mo- tion in-betweening,” ACM Transactions on Graphics (TOG) , vol. 39, no. 4, pp. 60–1, 2020

    2020

  29. [37]

    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 et al. , “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automation Letters , vol. 8, no. 6, pp. 3740–3747, 2023

    2023

  30. [38]

    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,” in Conference on Robot Learning . PMLR, 2022, pp. 91–100

    2022

  31. [39]

    Genesis: A universal and generative physics engine for robotics and beyond,

    G. Authors, “Genesis: A universal and generative physics engine for robotics and beyond,” December 2024. [Online]. Available: https://github.com/Genesis-Embodied-AI/Genesis

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.