Pith. sign in

REVIEW 3 major objections 6 minor 27 references

A single collision-aware planner with simulation in the loop turns one-camera human hand demos into executable robot-arm trajectories.

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

2026-07-13 02:25 UTC pith:IXYNFAQD

load-bearing objection Solid systems paper: multi-stage collision-aware retargeting works in sim and beats pose-by-pose baselines; the executable-trajectory claim is still sim-only. the 3 major comments →

arxiv 2607.09519 v1 pith:IXYNFAQD submitted 2026-07-10 cs.RO

DemoBridge: A Simulation-in-the-Loop Toolkit for Single-View Human Demonstration Retargeting

Zehao Wang , Fabien Despinoy , Sergey Zakharov , Tinne Tuytelaars , Rahaf Aljundi This is my paper
classification cs.RO
keywords human demonstration retargetingrobot manipulationcollision-aware trajectory optimizationsimulation-in-the-loopsingle-view hand trackingembodiment gapinverse kinematicsphysics validation
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.

DemoBridge claims that single-view RGB-stereo recordings of a human hand can be converted into robot-arm motions that are both faithful to the demonstration and physically valid. The hard part is the embodiment gap: a robot arm has far more collision volume than a hand, so pose-by-pose inverse kinematics often fails, and single-view tracking is noisy exactly at contact. The authors argue this is solved by one trajectory optimizer that jointly chooses grasps, avoids whole-arm and object collisions, and tracks the demonstrated path, with a physics simulator validating each phase and backtracking on failure. The result is a dynamically stable action sequence that can also serve as simulation data for policy learning, without manual grasp timing labels. If the method works as claimed, ordinary cameras become a practical source of robot manipulation data.

Core claim

A multi-stage collision-aware planner that optimizes an entire joint trajectory—jointly over alternative grasps, arm and grasped-object collision, and fidelity to the demonstrated end-effector path—plus simulation-in-the-loop validation and re-planning, converts noisy single-view human demonstrations into executable, physics-validated robot trajectories. On a synthetic cluttered benchmark it collides on at most one of fifty scenes while keeping endpoint error low and path fidelity high; on three real tabletop tasks it reaches placement on a majority of demos where pose-by-pose baselines largely fail.

What carries the argument

The multi-stage motion planner: a global route (demonstration sweep or RRT-Connect) seeds a trajectory optimizer that minimizes tracking, attachment, smoothness, collision, and limit costs in one nonlinear least-squares problem, specialized by masks per phase; a coordinator then runs each phase in a physics simulator and backtracks on failure.

Load-bearing premise

That object pose tracking and the grasp-release events derived from it stay accurate enough when the hand heavily occludes the object at contact, so the planner is not fed a systematically wrong reference.

What would settle it

Run the full pipeline on the same three real tasks but execute the retargeted trajectories on a physical robot arm and score reach-grasp-transport-place under real contact and gravity; success rates far below the simulated majority would falsify that the method produces truly executable motion from single-view demos.

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

If this is right

  • Ordinary single-camera human video can supply robot-executable, physics-validated trajectories without teleoperation, mocap suits, or multi-camera rigs.
  • Demonstrations that would be discarded as kinematically infeasible can be re-planned rather than lost, increasing usable data yield.
  • The same validated rollouts can seed real-to-sim-to-real and imitation pipelines as ready simulation trajectories.
  • Grasp and release timing need not be hand-annotated when co-motion and hand–object relation are used.
  • Perception backends, robot models, and pipeline stages can be swapped by configuration, so stronger trackers or new arms plug in without rewriting the retargeter.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If object tracking under occlusion remains the main bottleneck, coupling the planner with multi-view or contact-aware pose estimators would likely raise real-task place rates more than further planner tuning alone.
  • The modular coordinator-plus-planner design suggests a natural path to bimanual retargeting once dual-arm global routing and handover events are added, which the authors already flag as near-term work.
  • Policy learners trained only on these re-planned, sim-validated trajectories may inherit a bias toward collision-safe routes that diverge from the original human style when the demonstration was near the arm’s reach limit.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. DemoBridge is a modular toolkit that converts a single-view RGB-stereo recording of a human hand demonstration into a physics-validated robot-arm trajectory in Isaac Sim. The core contribution is a multi-stage, collision-aware retargeting planner that jointly optimizes the full joint trajectory over alternative grasps, whole-arm and grasped-object collision, and fidelity to the demonstrated end-effector path (Eqs. 1–5), with a simulation-in-the-loop coordinator that validates each phase and backtracks on failure. Grasp/release timing is inferred from kinematic co-motion rather than annotated. Perception backends, robot, and pipeline stages are configuration-swappable. Evaluation comprises (i) a controlled 50-scene synthetic planner benchmark isolating goal-only vs. with-reference settings and base vs. attach versions (Table I), and (ii) whole-pipeline retargeting on three real tabletop tasks with ten demonstrations each, compared against reimplemented pose-by-pose baselines (Table II).

Significance. If the results hold, the paper offers a practical, open-source bridge from passive single-view human video to robot-executable (or at least sim-executable) manipulation trajectories, addressing a real bottleneck for scaling imitation data without teleoperation or mocap. The multi-stage planner is a concrete technical contribution: the synthetic benchmark is carefully constructed with kinematic witness paths and controlled clutter, and the single-stage ablation cleanly isolates the global-routing stage (collisions drop from ~30/50 to ≤1/50 on base scenes). Code release, modular service architecture, and annotation-free event extraction are genuine strengths for a toolkit paper and support reuse for real-to-sim-to-real and policy-learning pipelines. The work is incremental relative to per-frame IK retargeting and trajectory optimization, but the joint treatment of noisy single-view references, whole-arm collision, grasp alternatives, and sim-in-the-loop backtracking is a useful systems-level advance.

major comments (3)
  1. [Abstract / Sec. VI-B / Table II] Abstract, Sec. I, and Sec. VI-B claim conversion into “executable,” “dynamically stable” robot-arm trajectories, yet all real-demonstration success (Table II) and the simulation-in-the-loop validation (Sec. V-B) are obtained entirely inside Isaac Sim. No physical robot execution, closed-loop sensing, or contact dynamics on hardware are reported. For a toolkit whose central promise is executable retargeting across the embodiment gap, this is a load-bearing gap: the synthetic benchmark (Table I) uses perfect geometry and cannot close sim-to-real transfer for perception residuals or contact. Either add hardware trials on at least one task, or substantially temper “executable” language to “sim-validated / ready-to-use simulation rollouts” throughout abstract, intro, and conclusions, and discuss the remaining transfer risk explicitly.
  2. [Sec. VII / Sec. V-A / Fig. 3C / Table II] Sec. VII correctly names per-frame 6D object tracking under hand occlusion as the main limitation, and Fig. 3C shows an inverted object orientation from tracking error. Nonetheless, the real-demo evaluation still treats the reconstructed digital twin and tracked poses as the reference fed to the planner and scores success only under that twin. Because event extraction (Sec. V-A) and attachment state depend on these tracks, systematic contact-time corruption would invalidate both the reference path and the grasp/release events. The paper should quantify tracking error near contact (e.g., pose drift / loss rate over the grasp window) and report how often event extraction or attachment state is wrong, or provide an ablation with perturbed/occluded tracks, so that Table II can be read as retargeting quality rather than joint perception+planning under optimistic geometry.
  3. [Sec. V-C Eq. (2) / Sec. VI-A / Table I] Table I and Sec. VI-A present a strong planner comparison, but free parameters that shape collision and seating behavior—clearance buffer δ, taper window τ (Eq. 2), sphere approximation (100 spheres of radius 4 mm), and the 5 mm / 2 mm success and collision thresholds—are not ablated or justified beyond construction convenience. Because multi-stage superiority is argued largely via near-zero collisions and endpoint accuracy under these choices, a short sensitivity study (or fixed public defaults with rationale) is needed to show that the global-routing benefit is not an artifact of a particular buffer/taper setting.
minor comments (6)
  1. [Abstract] Abstract and opening sentence: “an toolkit” → “a toolkit” (repeated grammar error).
  2. [Fig. 1 / Sec. III] Fig. 1 caption and Sec. III: the four-layer architecture is clear, but the figure text is dense; a short legend distinguishing Pipeline vs. Services vs. Data would help readers scanning the system overview.
  3. [Sec. VI-B / Table II] Sec. VI-B baselines: YOTO† and RoboWheel† are reimplementations with manually annotated grasp/release frames. State explicitly in the table caption or text that this gives the baselines oracle timing (an advantage to them), so readers do not misread the comparison as fully end-to-end.
  4. [Sec. V-C Eq. (1)] Eq. (1): define the binary mask vector m and the attachment/phase gating more formally (which terms are zeroed in rest vs. grasp vs. transport) so the “single objective covers all phases” claim is fully checkable.
  5. [Sec. VI-A Metrics] nDTW is reported without a precise formula or reference normalization; a one-line definition (or citation) would aid reproducibility of Table I path-fidelity scores.
  6. [Sec. II] Related work could briefly position against other whole-trajectory / optimization-based retargeting and collision-aware IK toolkits beyond RoboWheel, YOTO, SPIDER, and FlowHOI, to clarify novelty of the global-route + masked trajectory-opt combination.

Circularity Check

0 steps flagged

No significant circularity: systems toolkit with independent synthetic feasibility witnesses and external baseline reimplementations; planner objective is not tautological with reported metrics.

full rationale

DemoBridge is an engineering/systems paper whose load-bearing claims are empirical (Table I planner metrics; Table II cumulative stage success) rather than first-principles predictions. The multi-stage planner (Eqs. 1–5) minimizes a stated nonlinear least-squares cost (track, attach, smooth, coll, lim) seeded by a collision-aware reference sweep or RRT; success is scored by independent criteria (endpoint <5 mm, >2 mm penetration audit, nDTW). The synthetic benchmark deliberately admits a witness path by construction so that feasibility is known, but the planner is not given that witness as a free solution—goal-only must discover a route, and with-reference must track while remaining collision-free; the single-stage ablation’s 29/50 collisions shows the evaluation is not forced by the scene generator. Real-demo comparisons use reimplemented third-party recipes (YOTO†, RoboWheel†) with manually supplied grasp timing that the authors’ own pipeline does not receive. Overlapping-author perception backends (e.g. RecGen) are swappable config services, not uniqueness theorems or fitted constants renamed as predictions. No equation reduces a reported result to its own input by definition; residual self-citation is ordinary and non-load-bearing. Score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

DemoBridge is an engineering systems paper. Its central claims rest on standard motion-planning mathematics, ordinary robotics domain assumptions (kinematic co-motion, MANO anatomy, Isaac Sim as validator), and a modest set of hand-chosen planner hyperparameters and success thresholds. No new physical entities are postulated; the ‘invented’ pieces are software modules whose only evidence is the paper’s own experiments.

free parameters (4)
  • clearance buffer δ and taper window τ
    Hand-chosen safety margins that control how aggressively the optimizer approaches the grasp object versus true obstacles (Eq. 2); values are not derived from first principles.
  • object sphere approximation (100 spheres of radius 4 mm)
    Collision geometry discretization chosen to match the real pipeline; radius and count affect both feasibility and computational cost.
  • endpoint success threshold 5 mm and collision penetration threshold 2 mm
    Binary success criteria used for all synthetic metrics; different thresholds would change reported success rates.
  • Savitzky–Golay / median filter parameters and MANO fitting residual
    Noise-cleaning hyperparameters that determine the quality of the reference trajectory and contact events fed to the planner.
axioms (5)
  • domain assumption Kinematic co-motion of hand and object over a temporal window is a reliable indicator of grasp/hold state
    Core of the annotation-free event extractor (Sec. V-A); if co-motion is produced by external forces or tracking noise, grasp timing is wrong.
  • domain assumption A human demonstration path lies in a collision-free homotopy class usable as a global seed for the robot arm
    Justifies the ‘demo-given’ branch of Eq. 4; embodiment gap or severe clutter can place the human path in an infeasible class for the robot.
  • domain assumption Isaac Sim physics (contacts, gravity, gripper dynamics) is a sufficient validator of real-world executability
    All real-demo success metrics are measured inside the simulator (Sec. VI-B); no hardware closed-loop results are reported.
  • standard math Nonlinear least-squares trajectory optimization with augmented-Lagrangian joint-limit and stand-off constraints converges to a usable local minimum when seeded in the correct homotopy class
    Standard assumption underlying PyRoKi-based refinement (Eqs. 1 and 5).
  • standard math RRT-Connect returns a collision-free geometric path when one exists in the discretized configuration space
    Used for the no-reference global seed (Eq. 4).
invented entities (2)
  • phase-masked multi-stage retargeting planner (global route + trajectory optimizer with attachment/collision masks) no independent evidence
    purpose: Single objective that covers goal-only reach, reference following, and grasped-object transport while keeping sparsity pattern fixed across phases
    The specific combination of binary masks, dual world/gripper collision primitives, and sim-in-loop backtracking is introduced here; evidence is only the paper’s own synthetic and real-demo tables.
  • simulation-in-the-loop coordinator with automatic backtracking over ranked grasp candidates no independent evidence
    purpose: Validate each manipulation phase in Isaac Sim and re-plan rather than discard an infeasible demonstration
    Software control loop unique to DemoBridge; no external falsifiable prediction beyond the reported success rates.

pith-pipeline@v1.1.0-grok45 · 15958 in / 3500 out tokens · 50049 ms · 2026-07-13T02:25:45.145507+00:00 · methodology

0 comments
read the original abstract

We present DemoBridge, an toolkit that turns a single-view RGB stereo recording of a human hand demonstration into an executable, physics-validated robot-arm trajectory. Retargeting across the embodiment gap is hard. A robot arm reaches a target with a long, articulated body whose links carry far more collision volume than a hand. Solving inverse kinematics for the mapped end-effector pose often yields no collision-free solution, and a trajectory imposes this at every waypoint. A single view adds noise, leaving the demonstrated reference inaccurate. At the core of DemoBridge is a single collision-aware planner. It optimizes the whole joint trajectory at once, reasoning jointly over alternative grasp poses, whole-arm and grasped-object collision, and fidelity to the demonstrated path. A physics simulator runs in the loop. It validates each phase as it is produced and backtracks on failure, so a demonstration that cannot be reproduced as given is re-planned rather than discarded. The resulting action sequence is dynamically stable and faithful to the demonstrated manipulation. It also doubles as a ready-to-use simulation rollout for policy learning. Grasp timing is inferred automatically, and the perception backends, robot, and pipeline stages are swappable from configuration. We evaluate whole-pipeline retargeting on three real-demonstration tasks and the planner on a controlled synthetic benchmark. Our code is available at https://gitlab.kuleuven.be/u0123974/demo-bridge/ .

Figures

Figures reproduced from arXiv: 2607.09519 by Fabien Despinoy, Rahaf Aljundi, Sergey Zakharov, Tinne Tuytelaars, Zehao Wang.

Figure 1
Figure 1. Figure 1: DemoBridge’s four-layer architecture. A thin UI/CLI triggers [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example generated synthetic scenes. Left: a grasp (base) scene, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Real-world retargeting results from single-view human demonstrations. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 7 linked inside Pith

  1. [1]

    Dream to manipulate: Compositional world models empowering robot imitation learning with imagination, 2024

    Leonardo Barcellona, Andrii Zadaianchuk, Davide Allegro, Samuele Papa, Stefano Ghidoni, and Efstratios Gavves. Dream to manipulate: Compositional world models empowering robot imitation learning with imagination, 2024. URL https://arxiv. org/abs/2412.14957

    Pith/arXiv arXiv 2024

  2. [2]

    X-Sim: Cross-embodiment learning via real-to-sim- to-real.arXiv preprint arXiv:2505.07096, 2025

    Prithwish Dan, Kushal Kedia, Angela Chao, Edward Weiyi Duan, Maximus Adrian Pace, Wei-Chiu Ma, and Sanjiban Choudhury. X-Sim: Cross-embodiment learning via real-to-sim- to-real.arXiv preprint arXiv:2505.07096, 2025

    arXiv 2025

  3. [3]

    PyRoKi: A modular toolkit for robot kinematic optimization.arXiv preprint, 2025

    Chung Min Kim, Brent Yi, Hongsuk Choi, Yi Ma, Ken Gold- berg, and Angjoo Kanazawa. PyRoKi: A modular toolkit for robot kinematic optimization.arXiv preprint, 2025

    2025

  4. [4]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  5. [5]

    Kuffner and Steven M

    James J. Kuffner and Steven M. LaValle. RRT-Connect: An effi- cient approach to single-query path planning. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), volume 2, pages 995–1001, 2000

    2000

  6. [6]

    Any6D: Model-free 6D pose estimation of novel objects.arXiv preprint, 2025

    Taeyeop Lee et al. Any6D: Model-free 6D pose estimation of novel objects.arXiv preprint, 2025

    2025

  7. [7]

    A method for the solution of certain non-linear problems in least squares.Quarterly of Applied Mathematics, 2(2):164–168, 1944

    Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares.Quarterly of Applied Mathematics, 2(2):164–168, 1944

    1944

  8. [8]

    Santos, Brandon Rothrock, and Song- Chun Zhu

    Hangxin Liu, Xu Xie, Matt Millar, Mark Edmonds, Feng Gao, Yixin Zhu, Veronica J. Santos, Brandon Rothrock, and Song- Chun Zhu. A glove-based system for studying hand-object manipulation via joint pose and force sensing. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017

    2017

  9. [9]

    Tyler Ga Wei Lum, Olivia Y . Lee, C. Karen Liu, and Jeannette Bohg. Crossing the human-robot embodiment gap with sim- to-real rl using one human demonstration, 2025. URL https: //arxiv.org/abs/2504.12609

    Pith/arXiv arXiv 2025

  10. [10]

    SAM 3D: Segment anything in 3d.arXiv preprint, 2025

    Meta AI. SAM 3D: Segment anything in 3d.arXiv preprint, 2025

    2025

  11. [11]

    Isaac Lab: A GPU-accelerated simulation framework for robot learning

    NVIDIA. Isaac Lab: A GPU-accelerated simulation framework for robot learning. https://isaac-sim.github.io/IsaacLab, 2023

    2023

  12. [12]

    Spider: Scalable physics-informed dexterous retargeting, 2025

    Chaoyi Pan, Changhao Wang, Haozhi Qi, Zixi Liu, Homanga Bharadhwaj, Akash Sharma, Tingfan Wu, Guanya Shi, Jitendra Malik, and Francois Hogan. Spider: Scalable physics-informed dexterous retargeting, 2025. URL https://arxiv.org/abs/2511. 09484

    2025

  13. [13]

    Reconstruct- ing hands in 3D with transformers

    Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstruct- ing hands in 3D with transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2024

    2024

  14. [14]

    From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation

    Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. InIEEE Robotics and Au- tomation Letters (RA-L), 2022

    2022

  15. [15]

    Abraham Savitzky and Marcel J. E. Golay. Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36(8):1627–1639, 1964

    1964

  16. [16]

    cuRoboV2: Dynamics-aware motion generation with depth-fused distance fields for high-dof robots, 2026

    Balakumar Sundaralingam, Adithyavairavan Murali, and Stan Birchfield. cuRoboV2: Dynamics-aware motion generation with depth-fused distance fields for high-dof robots, 2026. URL https://arxiv.org/abs/2603.05493

    Pith/arXiv arXiv 2026

  17. [17]

    TRI-Stereo: Learned stereo depth

    Toyota Research Institute. TRI-Stereo: Learned stereo depth. https://github.com/TRI-ML, 2024

    2024

  18. [18]

    Karen Liu

    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. In Robotics: Science and Systems (RSS), 2024

    2024

  19. [19]

    HO-Cap: A capture system and dataset for 3D reconstruction and pose tracking of hand-object interaction

    Jikai Wang, Qifan Zhang, Yu-Wei Chao, Bowen Wen, Xiaohu Guo, and Yu Xiang. HO-Cap: A capture system and dataset for 3D reconstruction and pose tracking of hand-object interaction. arXiv preprint arXiv:2406.06843, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Foun- dationPose: Unified 6D pose estimation and tracking of novel objects

    Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foun- dationPose: Unified 6D pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  21. [21]

    FoundationStereo: Zero-shot stereo matching.arXiv preprint, 2025

    Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. FoundationStereo: Zero-shot stereo matching.arXiv preprint, 2025

    2025

  22. [22]

    Reconstruction by generation: 3D multi- object scene reconstruction from sparse observations.arXiv preprint arXiv:2604.27106, 2026

    Andrii Zadaianchuk, Leonardo Barcellona, Lennard Schuene- mann, Christian Gumbsch, Zehao Wang, Muhammad Zubair Irshad, Fabien Despinoy, Rahaf Aljundi, Stratis Gavves, and Sergey Zakharov. Reconstruction by generation: 3D multi- object scene reconstruction from sparse observations.arXiv preprint arXiv:2604.27106, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    FlowHOI: Flow-based semantics-grounded generation of hand-object interactions for dexterous robot manipulation, 2026

    Huajian Zeng, Lingyun Chen, Jiaqi Yang, Yuantai Zhang, Fan Shi, Peidong Liu, and Xingxing Zuo. FlowHOI: Flow-based semantics-grounded generation of hand-object interactions for dexterous robot manipulation, 2026. URL https://arxiv.org/abs/ 2602.13444

    arXiv 2026

  24. [24]

    Zhai, Kuo-Hao Zeng, Jiasen Lu, Ali Farhadi, Shenlong Wang, and Wei-Chiu Ma

    Albert J. Zhai, Kuo-Hao Zeng, Jiasen Lu, Ali Farhadi, Shenlong Wang, and Wei-Chiu Ma. Imitating what works: Simulation- filtered modular policy learning from human videos, 2026. URL https://arxiv.org/abs/2602.13197

    arXiv 2026

  25. [25]

    MediaPipe Hands: On-device real-time hand tracking.arXiv preprint arXiv:2006.10214, 2020

    Fan Zhang, Valentin Bazarevsky, Andrey Vakunov, Andrei Tkachenka, George Sung, Chuo-Ling Chang, and Matthias Grundmann. MediaPipe Hands: On-device real-time hand tracking.arXiv preprint arXiv:2006.10214, 2020

    Pith/arXiv arXiv 2006

  26. [26]

    Robowheel: A data engine from real-world human demonstrations for cross-embodiment robotic learning.arXiv preprint arXiv:2512.02729, 2025

    Yuhong Zhang, Zihan Gao, Shengpeng Li, Ling-Hao Chen, Kaisheng Liu, Runqing Cheng, Xiao Lin, Junjia Liu, Zhuoheng Li, Jingyi Feng, et al. Robowheel: A data engine from real-world human demonstrations for cross-embodiment robotic learning.arXiv preprint arXiv:2512.02729, 2025

    arXiv 2025

  27. [27]

    You only teach once: Learn one-shot bimanual robotic manipulation from video demonstrations

    Huayi Zhou, Ruixiang Wang, Yunxin Tai, Yueci Deng, Guiliang Liu, and Kui Jia. You only teach once: Learn one-shot bimanual robotic manipulation from video demonstrations. InRobotics: Science and Systems (RSS), 2025. Extended version YOTO++ to appear in IEEE TPAMI 2026; arXiv:2501.14208

    Pith/arXiv arXiv 2025