Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why

Andreas Krause; Chenhao Li; Marco Hutter

REVIEW 3 major objections 5 minor 26 references

Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why

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 Feature-based and GAN-based imitation learning have no universal winner.

desk verdict A competent, honest survey with a genuinely useful debunking section, but its central 'when and why' axis is partly confounded: the paper's own evidence shows that representation and alignment choices, not the explicit/implicit reward dichotomy, drive many of the differences it attributes to feature-based vs GAN-based methods. read the letter →

arxiv 2507.05906 v2 pith:HJ4WZDSF submitted 2025-07-08 cs.LG cs.AIcs.GRcs.RO

Chenhao Li , Marco Hutter , Andreas Krause This is my paper

classification cs.LGcs.AIcs.GRcs.RO

keywords imitationlearningfromdemonstrationsGAN-basedfeature-basedrewardsadversarialmotionpriorsstructuredrepresentationsphysics-basedcontrolrewarddesign

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

This paper is a survey that asks when a practitioner should use feature-based imitation, which scores an agent's behavior against hand-crafted motion features, and when a GAN-based approach, which learns an implicit reward from a discriminator, is the better choice. The authors argue that neither paradigm is universally superior: feature-based rewards give dense, interpretable supervision and high-fidelity tracking but can be brittle and hard to generalize, while adversarial rewards scale to diverse unstructured motion data but suffer from training instability, mode collapse, and coarse signal. The central practical claim is that method choice should be guided by task priorities, such as fidelity versus diversity, interpretability versus flexibility, and training simplicity versus large-scale generalization, rather than by paradigm loyalty. The survey also argues that recent work in both camps converges on structured motion representations, which make transitions smoother, behavior more controllable, and task integration easier. A sympathetic reader would take this as a framework for principled method selection and as a warning that common beliefs about GAN-based methods are often overstated.

What carries the argument

The central object is the reward signal's structure: explicit, feature-based rewards versus implicit, adversarially learned rewards. Feature-based machinery centers on a phase variable that time-aligns the policy with a reference clip and on hand-crafted features such as joint positions, velocities, and end-effector poses; GAN-based machinery centers on a discriminator that distinguishes policy-generated transitions from reference snippets and thereby defines an implicit similarity metric. A second, increasingly load-bearing mechanism is the structured motion representation, a learned or annotated latent embedding of motions that conditions the policy and reward, supports smooth interpolation between behaviors, and is what the paper identifies as the emerging convergence point between the paradigms.

What would settle it

A controlled benchmark that varies task priorities, such as fidelity, diversity, stability, and adaptability, on a common motion dataset and finds one paradigm winning across all priority settings would falsify the core claim that choice should be task-guided; equally, demonstrating that a feature-based method with a fixed representation matches GAN-based scalability on unstructured data without any adversarial component would challenge the convergence narrative.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the feature-based versus GAN-based distinction, when viewed through reward structure, is a spectrum rather than a dichotomy. Feature-based methods such as DeepMimic derive dense per-frame rewards from explicit features and a phase variable, producing high-fidelity imitation but limited multi-clip generalization and abrupt transitions. GAN-based methods such as AMP replace this with a discriminator that scores short state-transition windows, giving distributional supervision that scales to unstructured datasets but is prone to discriminator saturation and mode collapse. The paper's central thesis is that these failure modes mirror each other, with mode collapse in GANs matching the brittleness of poor motion representations in feature-based methods, and that the field is converging on structured motion representations as the bridge between the two. The conclusion is that selection should be driven by application constraints and priorities, not by a belief in universal superiority.

Load-bearing premise

The load-bearing premise is that the feature-based versus GAN-based split, defined over state-only offline demonstrations in physics-based control, is the right axis for choosing an imitation method; if the methods people actually use are action-labeled behavior cloning or diffusion policies, the survey's guidance does not cover the dominant practice.

Editorial extensions

If this is right

Method selection becomes a task-level design decision: high-fidelity reproduction favors feature-based rewards, while large-scale diversity favors adversarial or latent-conditioned approaches.
Claims that GAN-based methods need no hand-crafted features or reward weights do not hold, because discriminator inputs and feature scaling act as implicit feature choices.
Modern feature-based methods with learned motion embeddings can produce transitions as smooth as GAN-based ones, so smoothness alone does not justify choosing an adversarial method.
Scalability and generalization are properties of the motion representation, not of the reward paradigm; both camps need structured embeddings to scale to large datasets.
Evaluation metrics such as motion naturalness or energy efficiency mostly reflect the demonstration data and should not be used to compare algorithms; comparisons should target reward quality, stability, generalization, and task adaptability.

Reading between the lines

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

A natural extension is to re-run the comparison including action-labeled behavior cloning and diffusion policies, which the survey deliberately excludes; the reward-structure axis would need to be widened to explicit action supervision versus implicit state-based supervision.
One testable prediction is that replacing a GAN's discriminator with a learned feature-based distance updated online would recover much of adversarial methods' scalability while improving stability, since the survey's convergence argument implies the two paradigms differ less than their names suggest.
For practitioners, the convergence claim suggests a hybrid recipe: use dense feature rewards for task-critical body parts, an adversarial term for style and diversity, and a shared latent motion embedding for both; the paper does not propose this, but its framework makes it the obvious next design.
If structured representations are truly the deciding factor, then benchmark comparisons should report representation type and capacity alongside the imitation algorithm; otherwise apparent wins may be representation effects mislabeled as paradigm effects.

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.

Desk Editor's Note

read the letter

You should know this paper is a survey, not a methods paper. It has no new algorithms, equations, or measurements. What it does well: it organizes a messy literature, gives a clear taxonomy, and its 'debunking common beliefs' section (Section 10) is genuinely useful—it punctures several lazy claims like 'GANs scale better' and 'feature-based methods generalize worse' by pointing out that representation quality, not the reward structure, often drives these outcomes.

The soft spots are real but not fatal. The central 'when and why' axis—feature-based vs GAN-based reward structure—is partly confounded. The paper itself notes that GAN scalability comes from short, phase-free transition windows, and that modern feature-based methods with learned embeddings gain similar scalability and smoothness. That is, the observed differences often track representation and alignment choices rather than the explicit/implicit reward dichotomy. Section 10 concedes this, but Table 2 and the framing in Sections 8 and 11 still map the trade-offs onto the reward-signal axis. A careful reader can reconcile the two, but a practitioner skimming the tables could walk away with the very misconception the paper says it wants to correct. The stress-test note is right about this, though the paper deserves credit for partially self-correcting.

The citation pattern is self-heavy (FLD, DFM, CASSI, WASABI, ConsMimic, Motion Priors Reimagined), but those are legitimate published works in a small niche, so it's not disqualifying. The explicit exclusion of behavior cloning and diffusion policies is disclosed up front; it limits the scope more than it misleads. Still, any 'when and why' guidance that omits the currently dominant large-scale manipulation approaches is incomplete for practice.

Who is this for? Someone new to physics-based imitation learning who wants a map of the field and a set of caveats to carry into method selection. It reads well, hedges judiciously, and doesn't overclaim. No serious referee would reject it as incoherent; a revision could usefully sharpen the causal story by separating 'reward structure' from 'representation and alignment' as independent axes.

My take: give it a serious referee. It's not a breakthrough, but it's a competent, honest survey that deserves to be engaged with. I'd bring it to reading group for the debunking section alone.

Referee Report

3 major / 5 minor

Summary. This survey compares feature-based and GAN-based methods for learning from demonstrations in physics-based control, organizing the field by reward structure: explicit, hand-crafted rewards versus implicit, adversarially learned objectives. It argues that neither paradigm is universally superior, that method choice should depend on task priorities such as fidelity, diversity, interpretability, and scalability, and that recent work in both paradigms converges on the use of structured motion representations. The paper provides a taxonomic table, a qualitative comparison table, a discussion of evaluation metrics, and a list of debunked common beliefs.

Significance. If taken as a conceptual framework, the survey performs a useful service by countering paradigm partisanship and by explicitly hedging many claims, especially in Section 10. It correctly observes that high-level outcome metrics are often dataset-dependent (Section 9) and that features are not absent in GAN-based methods, since discriminators operate on selected state features. However, the central 'when and why' framework is weakened by a confound between reward structure and temporal-alignment/representation choices, and the comparative claims in Table 2 are asserted rather than supported by systematic evidence. The paper also introduces a clear scope limitation by excluding action-labeled behavior cloning and diffusion policies, which narrows the practical guidance. The survey is best viewed as a well-organized qualitative synthesis rather than a validated decision framework.

major comments (3)

[Sections 5-7, 11, Table 2] The central dichotomy is confounded with representation and alignment choices. Section 6 explains GAN-based scalability by 'short, fixed-length transition windows' and the removal of 'phase-based or time-indexed alignment,' which are properties of the sampling scheme, not of the adversarial objective. Section 5 attributes DeepMimic's scalability limits to the one-hot motion identifier and to the phase variable, not to the explicit reward itself. Section 7 then shows feature-based methods with learned structured representations gaining scalability and smoothness. Table 2 and Section 11 nevertheless map these differences onto 'reward signal' and the 'GAN-based vs feature-based' axis. A practitioner following the framework could select a GAN-based method for diversity when the actual cause is windowed, phase-free supervision, or select a feature-based method for interpretability while ignoring that its alignment mechanism is the true limiting factor. The paper should disentangle the axes (reward structure, alignment mechanism, representation quality) or explicitly argue why they are inseparable.
[Table 2, Sections 8 and 11] The comparative claims are presented as conclusions but are not supported by quantitative or systematically summarized evidence. For example, the claims that GAN-based methods have 'training stability challenging' and feature-based methods are 'stable but sensitive to inductive bias' are plausible mechanistic statements, but the survey does not report effect sizes, training curves, or any meta-analysis of the cited papers. Since the stated goal is 'principled decision-making,' the authors should either label these claims as qualitative hypotheses that require empirical testing or include a structured evidence table with reported results (e.g., success rates, tracking errors, diversity metrics) from the surveyed works.
[Section 1, Title, Abstract] The scope explicitly excludes action-labeled behavior cloning and diffusion policies (e.g., Diffusion Policy, Gr00t N1), which are currently dominant in large-scale manipulation and teleoperation. The title and abstract promise a general comparison of learning from demonstrations, but the guidance is limited to physics-based control with state-only reference data. This is a legitimate delimitation, but it should be reflected in the title or addressed by a discussion of how the feature/GAN framework transfers to these excluded paradigms; otherwise the 'when and why' answer is incomplete for a significant part of the current field.

minor comments (5)

[General typography and references] Several author names and terms contain spacing artifacts, such as 'V ollenweider', 'W ASABI', and 'JoÃG, o Pedro AraÃšjo' in the references; a careful proofread is needed.
[Figures 1-4] The notation '𝐼reference', 'state Ƹ𝑠', and the reward labels are not defined in the captions; a short explanation of the reward signals and state variables would improve readability.
[Table 1] Table 1 is a long flat list of methods; grouping entries by subcategory (e.g., latent-conditioned, classifier-based, representation-based) would align the table with the Section 6 and 7 narrative and make the taxonomy easier to use.
[Section 9] The discussion of metrics is valuable, but the claim that high-level metrics are 'not inherent to the learning algorithm' could be sharpened by noting that some metrics (e.g., tracking error) are direct optimization targets, while others (e.g., motion naturalness) genuinely depend on reference data quality.
[Section 10] The quotation formatting is inconsistent (curly versus straight quotes) and some bullet items are verbose; a light editorial pass would improve consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the survey's guidance is an interpretive synthesis, not a derivation, and its self-citations are illustrative, not load-bearing.

full rationale

This is a survey/position paper without a mathematical derivation chain. The central claims—that feature-based and GAN-based methods trade off fidelity/diversity/interpretability and that both families converge on structured motion representations—are literature-synthesis judgments, not predictions computed from fitted parameters or first-principles equations; hence they cannot be equivalent to their inputs by construction. The convergence thesis is supported by a broad evidence set that includes numerous works outside the authors' own group (e.g., ControlVAE, PhysicsVAE, PAE, ASE, CALM, MaskedMimic), while self-cited entries (CASSI, WASABI, FLD, DFM, ConsMimic, Motion Priors Reimagined) appear as examples within a taxonomy rather than as the sole justification for a forced conclusion. No uniqueness theorem is imported, no parameter is fit and then renamed as a prediction, and no category is defined in terms of the claim it is used to support. The reviewer's objection that scalability actually tracks windowed sampling and representation quality rather than the explicit/implicit reward axis is a substantive external-validity or omitted-variable concern, but it does not amount to circularity under the stated criteria: it does not show that the survey's recommendation is identical to its evidence by construction. Score 0.

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

This is a survey, so there are no equations derived, no fitted parameters, and no invented entities. The ledger records the framing assumptions that carry the qualitative conclusions: the scope definition, the MDP backdrop, the reward-structure organizing axis, and the structured-representation lens.

assumptions (4)

domain assumption Learning from demonstrations is defined as methods that use state-based offline reference data to derive a reward signal, explicitly excluding action-labeled behavior cloning (e.g., diffusion policy, Gr00t N1).
This scope definition (Section 1) carves out the survey's subject. If the boundary were drawn differently to include action-labeled BC and diffusion policies, the feature-based versus GAN-based dichotomy would be incomplete and the 'when and why' guidance would change.
standard math The MDP formulation with reward maximization is the correct formalism for physics-based control in character animation and robotics.
Section 3 introduces the MDP tuple; this is standard background, not an ad hoc assumption, but it frames the entire comparison.
domain assumption Reward structure (explicit feature-based versus implicit GAN-based) is the principal axis along which imitation algorithms should be compared.
The entire survey (Sections 5-7, Table 2) organizes methods by this criterion. It is a framing choice, not a proven fact, and it determines the structure of the conclusions.
domain assumption Structured motion representations are the key to scalability and generalization in both paradigms.
The survey's convergence thesis (Sections 6-8) asserts this using selected works; it is a qualitative synthesis rather than a verified universal law, and it underpins the recommendation to invest in representation learning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why." pith.science (2026). https://pith.science/paper/HJ4WZDSF

@misc{pith2026250705906,
  author       = {Pith},
  title        = {Pith review of: Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJ4WZDSF}},
  note         = {Machine review of arXiv:2507.05906}
}

read the original abstract

This survey provides a comparative analysis of feature-based and GAN-based approaches to learning from demonstrations, with a focus on the structure of reward functions and their implications for policy learning. Feature-based methods offer dense, interpretable rewards that excel at high-fidelity motion imitation, yet often require sophisticated representations of references and struggle with generalization in unstructured settings. GAN-based methods, in contrast, use implicit, distributional supervision that enables scalability and adaptation flexibility, but are prone to training instability and coarse reward signals. Recent advancements in both paradigms converge on the importance of structured motion representations, which enable smoother transitions, controllable synthesis, and improved task integration. We argue that the dichotomy between feature-based and GAN-based methods is increasingly nuanced: rather than one paradigm dominating the other, the choice should be guided by task-specific priorities such as fidelity, diversity, interpretability, and adaptability. This work outlines the algorithmic trade-offs and design considerations that underlie method selection, offering a framework for principled decision-making in learning from demonstrations.

Figures

Figures reproduced from arXiv: 2507.05906 by the authors.

**Figure 2.** GAN-based methods via adversarial rewards. A discriminator learns to distinguish short [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Latent-conditioned GAN-based methods. The policy and discriminator are jointly con [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Feature-based methods with structured motion representations. The policy receives per [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 7 canonical work pages

[1]

Gr00t n1: An open foundation model for generalist humanoid robots

Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734,

arXiv
[3]

Expressive whole-body control for humanoid robots

Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, and Xiaolong Wang. Expressive whole-body control for humanoid robots. arXiv preprint arXiv:2402.16796,

arXiv
[5]

Adversarial motion priors make good substitutes for complex reward functions

Alejandro Escontrela, Xue Bin Peng, Wenhao Yu, Tingnan Zhang, Atil Iscen, Ken Goldberg, and Pieter Abbeel. Adversarial motion priors make good substitutes for complex reward functions. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 25–32. IEEE,

2022
[7]

Humanplus: Humanoid shadowing and imitation from humans

Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454,

arXiv
[9]

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

Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024a. Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-...

arXiv
[11]

Versatile skill control via self-supervised adversarial imitation of unlabeled mixed motions

Chenhao Li, Sebastian Blaes, Pavel Kolev, Marin Vlastelica, Jonas Frey, and Georg Martius. Versatile skill control via self-supervised adversarial imitation of unlabeled mixed motions. In 2023 IEEE international conference on robotics and automation (ICRA), pp. 2944–2950. IEEE, 2023a. Chenhao Li, Marin Vlastelica, Sebastian Blaes, Jonas Frey, Felix Grimmi...

arXiv 2023
[12]

Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control

Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Ri-Zhao Qiu, and Xiaolong Wang. Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control. arXiv preprint arXiv:2505.03738,

arXiv
[13]

Perpetual humanoid control for real-time simulated avatars

Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10895–10904, 2023a. Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations f...

arXiv

Show all 26 references

[14]

Robot motion diffusion model: Motion generation for robotic characters

Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, and Moritz Bächer. Robot motion diffusion model: Motion generation for robotic characters. In SIGGRAPH Asia 2024 Conference Papers, pp. 1–9, 2024a. Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, and Moritz Bächer. ...

2024
[15]

Humanmimic: Learning natural locomotion and transitions for humanoid robot via wasserstein adversarial imitation

Annan Tang, Takuma Hiraoka, Naoki Hiraoka, Fan Shi, Kento Kawaharazuka, Kunio Kojima, Kei Okada, and Masayuki Inaba. Humanmimic: Learning natural locomotion and transitions for humanoid robot via wasserstein adversarial imitation. In 2024 IEEE International Conference on Robot...

2024
[16]

Gemini robotics: Bringing ai into the physical world

Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montser- rat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world. arXiv preprint arXiv:2...

arXiv
[17]

Calm: Conditional adversarial latent models for directable virtual characters

13 Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, pp. 1–9,

2023
[18]

Zero-shot whole-body humanoid control via behavioral foundation models

Andrea Tirinzoni, Ahmed Touati, Jesse Farebrother, Mateusz Guzek, Anssi Kanervisto, Yingchen Xu, Alessandro Lazaric, and Matteo Pirotta. Zero-shot whole-body humanoid control via behavioral foundation models. arXiv preprint arXiv:2504.11054,

arXiv
[19]

Advanced skills through multiple adversarial motion priors in reinforcement learning

Eric V ollenweider, Marko Bjelonic, Victor Klemm, Nikita Rudin, Joonho Lee, and Marco Hutter. Advanced skills through multiple adversarial motion priors in reinforcement learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 5120–5126. IEEE,

2023
[20]

Physhoi: Physics- based imitation of dynamic human-object interaction

Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics- based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393,

arXiv
[21]

Dfm: Deep fourier mimic for expressive dance motion learning

Ryo Watanabe, Chenhao Li, and Marco Hutter. Dfm: Deep fourier mimic for expressive dance motion learning. arXiv preprint arXiv:2502.10980,

arXiv
[22]

Constrained style learning from imperfect demonstrations under task optimality

Kehan Wen, Chenhao Li, Junzhe He, and Marco Hutter. Constrained style learning from imperfect demonstrations under task optimality. arXiv preprint arXiv:2507.09371,
[23]

Parc: Physics-based augmentation with reinforcement learning for character controllers

Michael Xu, Yi Shi, KangKang Yin, and Xue Bin Peng. Parc: Physics-based augmentation with reinforcement learning for character controllers. arXiv preprint arXiv:2505.04002,

arXiv
[24]

Robotkeyframing: Learning locomotion with high-level objectives via mixture of dense and sparse rewards

Fatemeh Zargarbashi, Jin Cheng, Dongho Kang, Robert Sumner, and Stelian Coros. Robotkeyframing: Learning locomotion with high-level objectives via mixture of dense and sparse rewards. arXiv preprint arXiv:2407.11562,

arXiv
[25]

Twist: Teleoperated whole-body imitation system

Yanjie Ze, Zixuan Chen, JoÃG, o Pedro AraÃšjo, Zi-ang Cao, Xue Bin Peng, Jiajun Wu, and C Karen Liu. Twist: Teleoperated whole-body imitation system. arXiv preprint arXiv:2505.02833,

arXiv
[26]

Motion priors reimagined: Adapting flat-terrain skills for complex quadruped mobility

Zewei Zhang, Chenhao Li, Takahiro Miki, and Marco Hutter. Motion priors reimagined: Adapting flat-terrain skills for complex quadruped mobility. arXiv preprint arXiv:2505.16084,

arXiv
[2016]

Exbody2: Advanced expressive humanoid whole-body control

Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196,

arXiv
[2022]

One policy but many worlds: A scalable unified policy for versatile humanoid locomotion

Yahao Fan, Tianxiang Gui, Kaiyang Ji, Shutong Ding, Chixuan Zhang, Jiayuan Gu, Jingyi Yu, Jingya Wang, and Ye Shi. One policy but many worlds: A scalable unified policy for versatile humanoid locomotion. arXiv preprint arXiv:2505.18780,

arXiv
[2023]

C · ase: Learning conditional adversarial skill embeddings for physics-based characters

Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. C · ase: Learning conditional adversarial skill embeddings for physics-based characters. In SIGGRAPH Asia 2023 Conference Papers, pp. 1–11,

2023
[2024]

Syn- thesizing physical character-scene interactions

Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Syn- thesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pp. 1–9,

2023
[2025]

Gmt: General motion tracking for humanoid whole-body control

Zixuan Chen, Mazeyu Ji, Xuxin Cheng, Xuanbin Peng, Xue Bin Peng, and Xiaolong Wang. Gmt: General motion tracking for humanoid whole-body control. arXiv preprint arXiv:2506.14770,

arXiv

Pith tools

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

[1] [1]

Gr00t n1: An open foundation model for generalist humanoid robots

Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734,

arXiv

[2] [3]

Expressive whole-body control for humanoid robots

Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, and Xiaolong Wang. Expressive whole-body control for humanoid robots. arXiv preprint arXiv:2402.16796,

arXiv

[3] [5]

Adversarial motion priors make good substitutes for complex reward functions

Alejandro Escontrela, Xue Bin Peng, Wenhao Yu, Tingnan Zhang, Atil Iscen, Ken Goldberg, and Pieter Abbeel. Adversarial motion priors make good substitutes for complex reward functions. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 25–32. IEEE,

2022

[4] [7]

Humanplus: Humanoid shadowing and imitation from humans

Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454,

arXiv

[5] [9]

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

Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024a. Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-...

arXiv

[6] [11]

Versatile skill control via self-supervised adversarial imitation of unlabeled mixed motions

Chenhao Li, Sebastian Blaes, Pavel Kolev, Marin Vlastelica, Jonas Frey, and Georg Martius. Versatile skill control via self-supervised adversarial imitation of unlabeled mixed motions. In 2023 IEEE international conference on robotics and automation (ICRA), pp. 2944–2950. IEEE, 2023a. Chenhao Li, Marin Vlastelica, Sebastian Blaes, Jonas Frey, Felix Grimmi...

arXiv 2023

[7] [12]

Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control

Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Ri-Zhao Qiu, and Xiaolong Wang. Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control. arXiv preprint arXiv:2505.03738,

arXiv

[8] [13]

Perpetual humanoid control for real-time simulated avatars

Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10895–10904, 2023a. Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations f...

arXiv