The Latent Color Subspace: Emergent Order in High-Dimensional Chaos

Jessica Bader; Mateusz Pach; Quentin Bouniot; Serge Belongie; Zeynep Akata

REVIEW 3 major objections 4 minor 34 references

Color in FLUX.1’s VAE latent space forms a structured Hue-Saturation-Lightness subspace that can be read out and edited with closed-form, training-free operations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 22:23 UTC pith:GP2KCISD

load-bearing objection We only have the LCS abstract; the attached full text is a different paper (HumDex), so the color-subspace claims cannot be audited. the 3 major comments →

arxiv 2603.12261 v2 pith:GP2KCISD submitted 2026-03-12 cs.LG cs.AIcs.CV

The Latent Color Subspace: Emergent Order in High-Dimensional Chaos

Mateusz Pach , Jessica Bader , Quentin Bouniot , Serge Belongie , Zeynep Akata This is my paper

classification cs.LG cs.AIcs.CV

keywords latent color subspaceVAE latent spaceFLUX.1text-to-imagecolor controlHSLtraining-free editingclosed-form manipulation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Modern text-to-image models still make fine color control hard because we poorly understand how color is stored inside their latent spaces. This paper shows that the Variational Autoencoder latent space of FLUX.1 [Dev] contains an emergent, low-dimensional Latent Color Subspace whose axes align with the familiar Hue, Saturation, and Lightness dimensions. Because the structure is explicit, the same geometry can both predict the color of a latent code and rewrite that color by simple algebraic edits—no extra training or optimization required. A sympathetic reader cares because the result turns an opaque high-dimensional representation into a transparent, editable color coordinate system that works at inference time.

Core claim

Inside the VAE latent space of FLUX.1 [Dev], color is not scattered chaotically; it organizes into a Latent Color Subspace whose principal directions correspond to Hue, Saturation, and Lightness. This geometric interpretation is strong enough to support both accurate color prediction from latents and explicit, training-free color control via closed-form latent-space arithmetic.

What carries the argument

The Latent Color Subspace (LCS): a low-dimensional linear structure discovered inside the FLUX VAE latent that aligns with HSL axes and thereby supplies the closed-form directions used for both color readout and color rewriting.

Load-bearing premise

The discovered latent directions truly act as causal Hue-Saturation-Lightness axes that remain valid across diverse content, lighting, and styles, rather than mere correlations on a narrow set of images.

What would settle it

Apply the claimed closed-form LCS edits to a held-out suite of latents spanning many object categories, lighting conditions, and artistic styles; if the resulting decoded images systematically fail to exhibit the intended hue, saturation, or lightness shifts while preserving identity, the LCS interpretation does not hold as a general control mechanism.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

We only have the LCS abstract; the attached full text is a different paper (HumDex), so the color-subspace claims cannot be audited.

read the letter

The punchline is simple: the target is arXiv 2603.12261 on a Latent Color Subspace in FLUX.1’s VAE, but the manuscript body we were given is HumDex (humanoid teleoperation, 2603.12260). So this is not a normal full-paper read. Everything below is constrained by that mismatch.

From the abstract alone, what is new is a model-specific claim: color in FLUX.1 [Dev]’s VAE latents is organized in a way that looks like Hue–Saturation–Lightness, and that structure is enough for both predicting color and editing it with closed-form, training-free latent ops. That sits in a long line of latent attribute directions (GAN walks, concept editing), but a clean HSL-like account plus a fully training-free control recipe for a widely used T2I VAE would still be useful engineering and a concrete interpretability result. Code is promised, which is the right kind of commitment.

What we cannot do is check soundness. There are no equations, discovery protocol, held-out prediction metrics, control baselines, ablations, or failure modes for LCS in the text we have. The soft spot the reader flags—correlational color axes vs causal, generalizable HSL control across content, lighting, and style—is exactly the right concern, and it is still open. Defining axes from color-correlated variation and then “predicting” along those same axes can go circular if validation is not cleanly held out; we simply cannot see whether they did that.

The HumDex body that was attached is a coherent robotics systems paper on its own terms (IMU teleop, learning-based hand retargeting, two-stage human→robot imitation, tables on collection and generalization). It is not evidence for LCS.

Who this is for: people working on controllable generation and VAE/latent interpretability for modern T2I models. Value is real if the full LCS paper and code match the abstract with quantitative, held-out control results. On the evidence in front of us, I would not cite it yet, and I would not put it in reading group until we have the actual manuscript. A serious editor should still send a complete LCS submission to referees rather than desk-reject on the abstract’s face—it is scoped, falsifiable, and practically relevant—but we cannot rubber-stamp soundness from what we were given.

Referee Report

3 major / 4 minor

Summary. The manuscript presents HumDex, a portable IMU-based teleoperation system for humanoid whole-body dexterous manipulation on a Unitree G1 with 20-DoF hands. It combines pelvis-centric GMR body retargeting, a lightweight MLP that maps five fingertip positions to hand joints, and a two-stage imitation pipeline (ACT) that pretrains on diverse human demonstrations then fine-tunes on robot teleoperation data. Experiments report higher collection efficiency and success than a vision/VR baseline (Table I), better hand-pose fidelity than optimization-based retargeting (Fig. 4, Table II), and improved OOD generalization on Pick Bread when human data is used (Table III). Code is promised open-source.

Significance. If the results hold, HumDex offers a practical, low-infrastructure route to high-quality whole-body dexterous data and a simple sequential recipe for transferring human motion priors across a large embodiment gap. Strengths include concrete hardware modularity (commercial and <$200 SlimeVR options), a training-free closed-form hand MLP after a short calibration set, and quantitative gains on long-horizon, bimanual, and articulated tasks that vision-based systems struggle with. The open-source commitment further raises potential impact for the humanoid community.

major comments (3)

Title/abstract mismatch: the supplied abstract and arXiv id claim a Latent Color Subspace (LCS) analysis of FLUX.1 VAE color, yet the full body is HumDex (humanoid teleoperation). No LCS equations, figures, or FLUX experiments appear. The central LCS claim is therefore unauditable from the manuscript as provided; either the correct LCS paper must be substituted or the HumDex abstract/title must be restored before any scientific evaluation of LCS can proceed.
Table III / §IV-C: the generalization claim rests on a single task (Pick Bread) with 50 robot episodes and human data that already covers the OOD axes. No multi-task transfer, no statistical significance, and no comparison to stronger co-training or domain-adaptation baselines are given. The Mix baseline’s 0% collapse is informative but does not by itself establish that sequential training is the only or best remedy for the embodiment gap.
Table I baseline: the vision/PICO+hand-tracking baseline fails Scan&Pack entirely due to occlusion. While this highlights an IMU advantage, the comparison confounds tracking modality with hand-control interface and operator workflow; a stronger optical or hybrid baseline (or an ablation that isolates occlusion) would be needed to quantify how much of the 26% time and 22.5 pp policy-success gains are attributable to IMU versus other system choices.

minor comments (4)

Eq. (1)–(2) and the ACT observation/action dimensions are stated inconsistently across body DoF counts (29/31/35); a single table of robot DoFs would remove ambiguity.
Fig. 4 qualitative poses lack quantitative metrics (e.g., fingertip error or contact success under open-loop replay).
Appendix A1 tracker-density ablation is useful but not referenced in the main text; a short pointer would help readers.
Several typos and formatting artifacts remain (e.g., “HumDex(Fig. 1)”, “T owel”, broken arXiv line breaks).

Circularity Check

0 steps flagged

No auditable circular derivation: supplied full text is HumDex (not LCS), and HumDex’s claims rest on empirical teleop/IL results rather than self-definitional math.

full rationale

The review target abstract (arXiv 2603.12261, Latent Color Subspace / FLUX VAE color) asserts that a discovered HSL-like latent structure both predicts and controls color via training-free closed-form edits, but the FULL MANUSCRIPT TEXT provided is a different paper (HumDex, humanoid teleoperation). No LCS equations, discovery protocol, held-out prediction setup, or control operators appear, so no load-bearing step of the LCS claim can be reduced by construction to its inputs. On the manuscript that is actually present, HumDex’s chain is: IMU/GMR body retargeting + MLP hand retargeting trained offline on optimization-paired fingertip→joint data; ACT policies trained on collected demos; optional two-stage human-pretrain then robot-finetune with proprioception approximated by previous action. Success rates (Tables I–III, A1) are external empirical checks against baselines and OOD settings, not fitted parameters renamed as predictions. Self-citations (e.g., authors’ prior VLA/manipulation works) are background, not uniqueness theorems that force the central result. Per hard rules, without a quotable reduction of a claimed prediction to its defining fit or a load-bearing self-citation chain, circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

With only the abstract, load-bearing premises are those required for any LCS-style claim: that FLUX’s VAE latents are a meaningful place to read/edit color; that color factors approximately linearly (or closed-form-editably) along HSL-like axes; and that edits transfer without retraining. No free parameters or invented physical entities appear in the abstract; LCS is an interpretive construct, not a new particle/force.

axioms (3)

domain assumption Semantic attributes such as color are encoded in FLUX.1’s VAE latent space in a form that admits closed-form, training-free manipulation.
Core premise of the abstract’s control claim; standard in latent-editing work but not guaranteed for this architecture.
ad hoc to paper The discovered latent structure meaningfully corresponds to human HSL axes (Hue, Saturation, Lightness), not merely to some other color-correlated basis.
The paper’s interpretive label ‘reflecting HSL’ is a modeling choice that must be validated against perceptual/colorimetric measures.
domain assumption Closed-form latent edits that change color leave other semantics sufficiently intact for practical control.
Implicit success criterion for ‘explicitly control color’ without stating entanglement bounds.

invented entities (1)

Latent Color Subspace (LCS) no independent evidence
purpose: Name the claimed HSL-like structure in FLUX.1 VAE latents used for color prediction and training-free control.
Interpretive construct introduced by the paper; independent evidence would be quantitative prediction/control results and released code, not available in the abstract alone.

pith-pipeline@v1.1.0-grok45 · 20528 in / 2631 out tokens · 26272 ms · 2026-07-14T22:23:11.060538+00:00 · methodology

0 comments

read the original abstract

Text-to-image generation models have advanced rapidly, yet achieving fine-grained control over generated images remains difficult, largely due to limited understanding of how semantic information is encoded. We develop an interpretation of the color representation in the Variational Autoencoder latent space of FLUX.1 [Dev], revealing a structure reflecting Hue, Saturation, and Lightness. We verify our Latent Color Subspace (LCS) interpretation by demonstrating that it can both predict and explicitly control color, introducing a fully training-free method in FLUX based solely on closed-form latent-space manipulation. Code is available at https://github.com/ExplainableML/LCS.

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 16 linked inside Pith

[1]

Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252, 2025

Joao Pedro Araujo, Yanjie Ze, Pei Xu, Jiajun Wu, and C Karen Liu. Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252, 2025

arXiv 2025
[2]

Homie: Humanoid loco- manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013, 2025

Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. Homie: Humanoid loco- manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013, 2025

Pith/arXiv arXiv 2025
[3]

Wujihand retargeting,

Guanqi He and Wentao Zhang. Wujihand retargeting,
[4]

* Equal contribution

URL https://github.com/wuji-technology/wuji retargeting. * Equal contribution
[5]

Vitacformer: Learning cross- modal representation for visuo-tactile dexterous manipu- lation, 2025

Liang Heng, Haoran Geng, Kaifeng Zhang, Pieter Abbeel, and Jitendra Malik. Vitacformer: Learning cross- modal representation for visuo-tactile dexterous manipu- lation, 2025. URL https://arxiv.org/abs/2506.15953

Pith/arXiv arXiv 2025
[6]

Rwor: Generating robot demonstrations from human hand collection for policy learning without robot

Liang Heng, Xiaoqi Li, Shangqing Mao, Jiaming Liu, Ruolin Liu, Jingli Wei, Yu-Kai Wang, Yueru Jia, Chenyang Gu, Rui Zhao, et al. Rwor: Generating robot demonstrations from human hand collection for policy learning without robot. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13544–13551. IEEE, 2025

2025
[7]

Imagine2act: Leveraging object-action motion consistency from imagined goals for robotic manipula- tion.arXiv preprint arXiv:2509.17125, 2025

Liang Heng, Jiadong Xu, Yiwen Wang, Xiaoqi Li, Muhe Cai, Yan Shen, Juan Zhu, Guanghui Ren, and Hao Dong. Imagine2act: Leveraging object-action motion consistency from imagined goals for robotic manipula- tion.arXiv preprint arXiv:2509.17125, 2025

Pith/arXiv arXiv 2025
[8]

Egomimic: Scaling imitation learning via egocentric video

Simar Kareer, Dhruv Patel, Ryan Punamiya, Pranay Mathur, Shuo Cheng, Chen Wang, Judy Hoffman, and Danfei Xu. Egomimic: Scaling imitation learning via egocentric video. In2025 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 13226– 13233. IEEE, 2025

2025
[9]

A self-correcting vision-language-action model for fast and slow system manipulation, 2025

Chenxuan Li, Jiaming Liu, Guanqun Wang, Xiaoqi Li, Sixiang Chen, Liang Heng, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, Kaichen Zhou, and Shanghang Zhang. A self-correcting vision-language-action model for fast and slow system manipulation, 2025. URL https://arxiv.org/ abs/2405.17418

Pith/arXiv arXiv 2025
[10]

Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control.Robotics: Science and Systems 2025, 2025

Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Rizhao Qiu, and Xiaolong Wang. Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control.Robotics: Science and Systems 2025, 2025

2025
[11]

3ds-vla: A 3d spatial- aware vision language action model for robust multi- task manipulation

Xiaoqi Li, Liang Heng, Jiaming Liu, Yan Shen, Chenyang Gu, Zhuoyang Liu, Hao Chen, Nuowei Han, Renrui Zhang, Hao Tang, et al. 3ds-vla: A 3d spatial- aware vision language action model for robust multi- task manipulation. In9th Annual Conference on Robot Learning
[12]

Object-centric prompt-driven vision-language- action model for robotic manipulation

Xiaoqi Li, Jingyun Xu, Mingxu Zhang, Jiaming Liu, Yan Shen, Iaroslav Ponomarenko, Jiahui Xu, Liang Heng, Siyuan Huang, Shanghang Zhang, and Hao Dong. Object-centric prompt-driven vision-language- action model for robotic manipulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27638–27648, June 2025

2025
[13]

Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks.arXiv preprint arXiv:2506.08931, 2025

Yixuan Li, Yutang Lin, Jieming Cui, Tengyu Liu, Wei Liang, Yixin Zhu, and Siyuan Huang. Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks.arXiv preprint arXiv:2506.08931, 2025

Pith/arXiv arXiv 2025
[14]

Dexflow: A unified approach for dexterous hand pose retargeting and interaction, 2025

Xiaoyi Lin, Kunpeng Yao, Lixin Xu, Xueqiang Wang, Xuetao Li, Yuchen Wang, and Miao Li. Dexflow: A unified approach for dexterous hand pose retargeting and interaction, 2025. URL https://arxiv.org/abs/2505.01083

Pith/arXiv arXiv 2025
[15]

Sonic: Supersizing motion tracking for natural humanoid whole-body control.arXiv preprint arXiv:2511.07820, 2025

Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Sirui Chen, Fernando Casta ˜neda, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, Xingye Da, Runyu Ding, Cyrus Hogg, Lina Song, Edy Lim, Eugene Jeong, Tairan He, Haoru Xue, Wenli Xiao, Zi Wang, Simon Yuen, Jan Kautz, Yan Chang, Umar Iqbal, Linxi Fan, and Yuke Zhu. Sonic: Supersizing motion tracking for natu...

Pith/arXiv arXiv 2025
[16]

Dexmachina: Func- tional retargeting for bimanual dexterous manipulation,

Zhao Mandi, Yifan Hou, Dieter Fox, Yashraj Narang, Ajay Mandlekar, and Shuran Song. Dexmachina: Func- tional retargeting for bimanual dexterous manipulation,
[17]

URL https://arxiv.org/abs/2505.24853

Pith/arXiv arXiv
[18]

R3m: A universal visual representation for robot manipulation.arXiv preprint arXiv:2203.12601, 2022

Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation.arXiv preprint arXiv:2203.12601, 2022

Pith/arXiv arXiv 2022
[19]

Humanoid policy˜ human policy.arXiv preprint arXiv:2503.13441, 2025

Ri-Zhao Qiu, Shiqi Yang, Xuxin Cheng, Chaitanya Chawla, Jialong Li, Tairan He, Ge Yan, David J Yoon, Ryan Hoque, Lars Paulsen, et al. Humanoid policy˜ human policy.arXiv preprint arXiv:2503.13441, 2025

arXiv 2025
[20]

Real-world robot learning with masked visual pre-training

Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, pages 416–426. PMLR, 2023

2023
[21]

Slimevr: Open source full- body tracking, 2025

SlimeVR Contributors. Slimevr: Open source full- body tracking, 2025. URL https://github.com/SlimeVR/ SlimeVR-Server. Open-source full-body tracking server for virtual reality

2025
[22]

Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei, and C Karen Liu. Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

Pith/arXiv arXiv 2024
[23]

Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators

Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, and Pieter Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163, 2024. doi: 10. 1109/IROS58592.2024.10801581

arXiv 2024
[24]

Masked visual pre-training for motor control

Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022

Pith/arXiv arXiv 2022
[25]

Analyzing key objectives in human-to-robot retargeting for dexterous manipulation,

Chendong Xin, Mingrui Yu, Yongpeng Jiang, Zhefeng Zhang, and Xiang Li. Analyzing key objectives in human-to-robot retargeting for dexterous manipulation,
[26]

URL https://arxiv.org/abs/2506.09384

arXiv
[27]

Geometric retargeting: A principled, ultrafast neural hand retargeting algorithm, 2025

Zhao-Heng Yin, Changhao Wang, Luis Pineda, Krishna Bodduluri, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. Geometric retargeting: A principled, ultrafast neural hand retargeting algorithm, 2025. URL https: //arxiv.org/abs/2503.07541

Pith/arXiv arXiv 2025
[28]

Dexteritygen: Foundation controller for unprecedented dexterity, 2025

Zhao-Heng Yin, Changhao Wang, Luis Pineda, Fran- cois Hogan, Krishna Bodduluri, Akash Sharma, Patrick Lancaster, Ishita Prasad, Mrinal Kalakrishnan, Jitendra Malik, Mike Lambeta, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. Dexteritygen: Foundation controller for unprecedented dexterity, 2025. URL https://arxiv.org/ abs/2502.04307

Pith/arXiv arXiv 2025
[29]

Karen Liu

Yanjie Ze, Zixuan Chen, Jo ˜ao Pedro Ara ´ujo, Zi ang Cao, Xue Bin Peng, Jiajun Wu, and C. Karen Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833, 2025

Pith/arXiv arXiv 2025
[30]

Karen Liu

Yanjie Ze, Siheng Zhao, Weizhuo Wang, Angjoo Kanazawa, Rocky Duan, Pieter Abbeel, Guanya Shi, Jiajun Wu, and C. Karen Liu. Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832, 2025

arXiv 2025
[31]

Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation, 2025

Rongyu Zhang, Menghang Dong, Yuan Zhang, Liang Heng, Xiaowei Chi, Gaole Dai, Li Du, Yuan Du, and Shanghang Zhang. Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation, 2025. URL https://arxiv. org/abs/2503.20384

Pith/arXiv arXiv 2025
[32]

Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manip- ulation with low-cost hardware, 2023. URL https://arxiv. org/abs/2304.13705. APPENDIX METHODADDITIONALDETAILS A. Hardware Setup We visualize our hardware integration and teleoperation setup in Fig. A1. Our system is composed of three key components: Fig. A1:...

Pith/arXiv arXiv 2023
[33]

To emulate the robot’s egocentric view, an Intel RealSense D435i camera is mounted on a neckband worn by the operator

Human Data Collection •Setup & Vision:The operator wears the VIRDYN suit and gloves without active robot execution. To emulate the robot’s egocentric view, an Intel RealSense D435i camera is mounted on a neckband worn by the operator. •Pipeline:Raw motion capture data is processed via the General Motion Retargeting (GMR) model. This optimization-based sol...
[34]

We utilize the robot’s built-in Intel RealSense D435i head camera for vision

Robot Data Collection •Setup & Vision:The operator controls the physical G1 robot via the teleoperation system. We utilize the robot’s built-in Intel RealSense D435i head camera for vision. •Pipeline:The operator’s motion is mapped to robot commands via the same GMR solver in real-time. These commands are executed by the robot and simultaneously fetched f...

[1] [1]

Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252, 2025

Joao Pedro Araujo, Yanjie Ze, Pei Xu, Jiajun Wu, and C Karen Liu. Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252, 2025

arXiv 2025

[2] [2]

Homie: Humanoid loco- manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013, 2025

Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. Homie: Humanoid loco- manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013, 2025

Pith/arXiv arXiv 2025

[3] [3]

Wujihand retargeting,

Guanqi He and Wentao Zhang. Wujihand retargeting,

[4] [4]

* Equal contribution

URL https://github.com/wuji-technology/wuji retargeting. * Equal contribution

[5] [5]

Vitacformer: Learning cross- modal representation for visuo-tactile dexterous manipu- lation, 2025

Liang Heng, Haoran Geng, Kaifeng Zhang, Pieter Abbeel, and Jitendra Malik. Vitacformer: Learning cross- modal representation for visuo-tactile dexterous manipu- lation, 2025. URL https://arxiv.org/abs/2506.15953

Pith/arXiv arXiv 2025

[6] [6]

Rwor: Generating robot demonstrations from human hand collection for policy learning without robot

Liang Heng, Xiaoqi Li, Shangqing Mao, Jiaming Liu, Ruolin Liu, Jingli Wei, Yu-Kai Wang, Yueru Jia, Chenyang Gu, Rui Zhao, et al. Rwor: Generating robot demonstrations from human hand collection for policy learning without robot. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13544–13551. IEEE, 2025

2025

[7] [7]

Imagine2act: Leveraging object-action motion consistency from imagined goals for robotic manipula- tion.arXiv preprint arXiv:2509.17125, 2025

Liang Heng, Jiadong Xu, Yiwen Wang, Xiaoqi Li, Muhe Cai, Yan Shen, Juan Zhu, Guanghui Ren, and Hao Dong. Imagine2act: Leveraging object-action motion consistency from imagined goals for robotic manipula- tion.arXiv preprint arXiv:2509.17125, 2025

Pith/arXiv arXiv 2025

[8] [8]

Egomimic: Scaling imitation learning via egocentric video

Simar Kareer, Dhruv Patel, Ryan Punamiya, Pranay Mathur, Shuo Cheng, Chen Wang, Judy Hoffman, and Danfei Xu. Egomimic: Scaling imitation learning via egocentric video. In2025 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 13226– 13233. IEEE, 2025

2025

[9] [9]

A self-correcting vision-language-action model for fast and slow system manipulation, 2025

Chenxuan Li, Jiaming Liu, Guanqun Wang, Xiaoqi Li, Sixiang Chen, Liang Heng, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, Kaichen Zhou, and Shanghang Zhang. A self-correcting vision-language-action model for fast and slow system manipulation, 2025. URL https://arxiv.org/ abs/2405.17418

Pith/arXiv arXiv 2025

[10] [10]

Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control.Robotics: Science and Systems 2025, 2025

Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Rizhao Qiu, and Xiaolong Wang. Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control.Robotics: Science and Systems 2025, 2025

2025

[11] [11]

3ds-vla: A 3d spatial- aware vision language action model for robust multi- task manipulation

Xiaoqi Li, Liang Heng, Jiaming Liu, Yan Shen, Chenyang Gu, Zhuoyang Liu, Hao Chen, Nuowei Han, Renrui Zhang, Hao Tang, et al. 3ds-vla: A 3d spatial- aware vision language action model for robust multi- task manipulation. In9th Annual Conference on Robot Learning

[12] [12]

Object-centric prompt-driven vision-language- action model for robotic manipulation

Xiaoqi Li, Jingyun Xu, Mingxu Zhang, Jiaming Liu, Yan Shen, Iaroslav Ponomarenko, Jiahui Xu, Liang Heng, Siyuan Huang, Shanghang Zhang, and Hao Dong. Object-centric prompt-driven vision-language- action model for robotic manipulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27638–27648, June 2025

2025

[13] [13]

Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks.arXiv preprint arXiv:2506.08931, 2025

Yixuan Li, Yutang Lin, Jieming Cui, Tengyu Liu, Wei Liang, Yixin Zhu, and Siyuan Huang. Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks.arXiv preprint arXiv:2506.08931, 2025

Pith/arXiv arXiv 2025

[14] [14]

Dexflow: A unified approach for dexterous hand pose retargeting and interaction, 2025

Xiaoyi Lin, Kunpeng Yao, Lixin Xu, Xueqiang Wang, Xuetao Li, Yuchen Wang, and Miao Li. Dexflow: A unified approach for dexterous hand pose retargeting and interaction, 2025. URL https://arxiv.org/abs/2505.01083

Pith/arXiv arXiv 2025

[15] [15]

Sonic: Supersizing motion tracking for natural humanoid whole-body control.arXiv preprint arXiv:2511.07820, 2025

Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Sirui Chen, Fernando Casta ˜neda, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, Xingye Da, Runyu Ding, Cyrus Hogg, Lina Song, Edy Lim, Eugene Jeong, Tairan He, Haoru Xue, Wenli Xiao, Zi Wang, Simon Yuen, Jan Kautz, Yan Chang, Umar Iqbal, Linxi Fan, and Yuke Zhu. Sonic: Supersizing motion tracking for natu...

Pith/arXiv arXiv 2025

[16] [16]

Dexmachina: Func- tional retargeting for bimanual dexterous manipulation,

Zhao Mandi, Yifan Hou, Dieter Fox, Yashraj Narang, Ajay Mandlekar, and Shuran Song. Dexmachina: Func- tional retargeting for bimanual dexterous manipulation,

[17] [17]

URL https://arxiv.org/abs/2505.24853

Pith/arXiv arXiv

[18] [18]

R3m: A universal visual representation for robot manipulation.arXiv preprint arXiv:2203.12601, 2022

Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation.arXiv preprint arXiv:2203.12601, 2022

Pith/arXiv arXiv 2022

[19] [19]

Humanoid policy˜ human policy.arXiv preprint arXiv:2503.13441, 2025

Ri-Zhao Qiu, Shiqi Yang, Xuxin Cheng, Chaitanya Chawla, Jialong Li, Tairan He, Ge Yan, David J Yoon, Ryan Hoque, Lars Paulsen, et al. Humanoid policy˜ human policy.arXiv preprint arXiv:2503.13441, 2025

arXiv 2025

[20] [20]

Real-world robot learning with masked visual pre-training

Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, pages 416–426. PMLR, 2023

2023

[21] [21]

Slimevr: Open source full- body tracking, 2025

SlimeVR Contributors. Slimevr: Open source full- body tracking, 2025. URL https://github.com/SlimeVR/ SlimeVR-Server. Open-source full-body tracking server for virtual reality

2025

[22] [22]

Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei, and C Karen Liu. Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

Pith/arXiv arXiv 2024

[23] [23]

Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators

Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, and Pieter Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163, 2024. doi: 10. 1109/IROS58592.2024.10801581

arXiv 2024

[24] [24]

Masked visual pre-training for motor control

Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022

Pith/arXiv arXiv 2022

[25] [25]

Analyzing key objectives in human-to-robot retargeting for dexterous manipulation,

Chendong Xin, Mingrui Yu, Yongpeng Jiang, Zhefeng Zhang, and Xiang Li. Analyzing key objectives in human-to-robot retargeting for dexterous manipulation,

[26] [26]

URL https://arxiv.org/abs/2506.09384

arXiv

[27] [27]

Geometric retargeting: A principled, ultrafast neural hand retargeting algorithm, 2025

Zhao-Heng Yin, Changhao Wang, Luis Pineda, Krishna Bodduluri, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. Geometric retargeting: A principled, ultrafast neural hand retargeting algorithm, 2025. URL https: //arxiv.org/abs/2503.07541

Pith/arXiv arXiv 2025

[28] [28]

Dexteritygen: Foundation controller for unprecedented dexterity, 2025

Zhao-Heng Yin, Changhao Wang, Luis Pineda, Fran- cois Hogan, Krishna Bodduluri, Akash Sharma, Patrick Lancaster, Ishita Prasad, Mrinal Kalakrishnan, Jitendra Malik, Mike Lambeta, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. Dexteritygen: Foundation controller for unprecedented dexterity, 2025. URL https://arxiv.org/ abs/2502.04307

Pith/arXiv arXiv 2025

[29] [29]

Karen Liu

Yanjie Ze, Zixuan Chen, Jo ˜ao Pedro Ara ´ujo, Zi ang Cao, Xue Bin Peng, Jiajun Wu, and C. Karen Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833, 2025

Pith/arXiv arXiv 2025

[30] [30]

Karen Liu

Yanjie Ze, Siheng Zhao, Weizhuo Wang, Angjoo Kanazawa, Rocky Duan, Pieter Abbeel, Guanya Shi, Jiajun Wu, and C. Karen Liu. Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832, 2025

arXiv 2025

[31] [31]

Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation, 2025

Rongyu Zhang, Menghang Dong, Yuan Zhang, Liang Heng, Xiaowei Chi, Gaole Dai, Li Du, Yuan Du, and Shanghang Zhang. Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation, 2025. URL https://arxiv. org/abs/2503.20384

Pith/arXiv arXiv 2025

[32] [32]

Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manip- ulation with low-cost hardware, 2023. URL https://arxiv. org/abs/2304.13705. APPENDIX METHODADDITIONALDETAILS A. Hardware Setup We visualize our hardware integration and teleoperation setup in Fig. A1. Our system is composed of three key components: Fig. A1:...

Pith/arXiv arXiv 2023

[33] [33]

To emulate the robot’s egocentric view, an Intel RealSense D435i camera is mounted on a neckband worn by the operator

Human Data Collection •Setup & Vision:The operator wears the VIRDYN suit and gloves without active robot execution. To emulate the robot’s egocentric view, an Intel RealSense D435i camera is mounted on a neckband worn by the operator. •Pipeline:Raw motion capture data is processed via the General Motion Retargeting (GMR) model. This optimization-based sol...

[34] [34]

We utilize the robot’s built-in Intel RealSense D435i head camera for vision

Robot Data Collection •Setup & Vision:The operator controls the physical G1 robot via the teleoperation system. We utilize the robot’s built-in Intel RealSense D435i head camera for vision. •Pipeline:The operator’s motion is mapped to robot commands via the same GMR solver in real-time. These commands are executed by the robot and simultaneously fetched f...