Pith. sign in

REVIEW 4 major objections 6 minor 42 references

DIRIGENt: End-To-End Robotic Imitation of Human Demonstrations Based on a Diffusion Model

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A diffusion model maps a single RGB frame of a human pose directly to a robot's joint values, letting the robot imitate without any pre-defined human-to-robot mapping.

desk verdict A useful diffusion-based imitator with a clever reversed-imitation dataset, but the headline claim about imitating arbitrary humans is not yet supported; the held-out folds show much larger errors and the test set is still robot-imitation, not natural motion. read the letter →

arxiv 2501.16800 v1 pith:UKFBP2G4 submitted 2025-01-28 cs.RO cs.AI

Josua Spisak , Matthias Kerzel , Stefan Wermter This is my paper · ORCID
classification cs.ROcs.AI
keywords roboticimitationdiffusionmodelhuman-robotcorrespondencejointvaluegenerationlearningfromdemonstrationRGBimageconditioninversekinematicshumanoidrobot
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces DIRIGENt, an end-to-end diffusion model that takes a single RGB image of a human demonstration and directly outputs the 26 joint values of a humanoid robot's two arms. To train it, the authors built the DIRI dataset by reversing the usual protocol: humans imitate the robot, so each recorded human pose is naturally paired with the robot pose it should reproduce. The model conditions on the image and, during training, also receives a noisy version of the desired joint values, which helps it avoid redundant joint configurations. On the dataset's random 90:10 split, the generated end-effector positions are on average 2 mm, 3 mm, and 2 mm from the demonstrated positions along the x, y, and z axes, and the full model outperforms the baseline HoRoPose. The intended payoff is that a robot can be taught new arm movements by simple demonstration rather than by manual programming or pre-defined human-to-robot mappings.

What carries the argument

The load-bearing mechanism is a diffusion-conditioned U-net with two contracting paths and one expanding path. One contracting path encodes the condition, a 256 by 256 RGB image of the human demonstrator; the other encodes the noisy target, which during training is the desired 26-dimensional joint vector with Gaussian noise added according to a 1000-level cosine schedule. Both streams are fused at multiple resolutions, and the expanding path decodes to the joint values, which then pass through a differentiable kinematic chain that yields the end-effector position. The loss is $L = \omega_1 \cdot MSE(D_{DO}, D_{TI}) + \omega_2 \cdot MSE(D_{DO}^{(eef)}, D_{TI}^{(eef)})$, so the model is trained to match both the full joint configuration and the Cartesian end-effector position of the original robot demonstration. At inference the noisy input is pure noise, and because the network is trained to predict the target output directly rather than the noise, a single forward pass suffices.

What would settle it

Record the same arm movements with humans demonstrating naturally instead of imitating the robot, and compare DIRIGENt's end-effector error on those frames with the reported 2 to 3 mm on the DIRI test split; if the error rises to the scale of the fold-10 cross-validation error (several centimetres) or worse, the reversed-imitation protocol is the source of the reported accuracy. A simpler check is to run the released model on a held-out video where the demonstrator moves at natural speed and with self-chosen trajectories.

Watch

Extended reading notes

Core claim

The central claim is that robotic imitation of human arm movements can be cast as a single learned map from an RGB frame to joint values, and that a diffusion model trained on reversed-imitation pairs learns this map better than a CNN-based pose-estimation pipeline. The authors argue that three design choices jointly produce this result: natural human-robot pose pairs from the DIRI dataset, a diffusion input that shows the model a noisy copy of the target joint vector during training, and an end-to-end perception-to-action architecture in which the condition is the raw RGB image rather than an estimated pose. With these choices, the model reaches average end-effector distances of 2 mm, 3 mm, and 2 mm on the random split, and when tested on entirely unseen participants in 10-fold cross-validation it keeps average distances of 4.8 cm, 6.1 cm, and 4.3 cm while still reproducing movement patterns. Ablations show that removing the noisy diffusion input roughly triples the axis errors, that RGB images outperform both 2D and 3D pose estimates as conditions, and that the combined joint and Cartesian losses beat either loss alone.

Load-bearing premise

The load-bearing premise is that having humans imitate a slow, predictable robot produces human-robot pose pairs that are representative of ordinary human demonstrations, so that learning from those pairs transfers to natural demonstrations.

Editorial extensions

If this is right

  • A robot trained this way can map a novel human demonstrator's arm pose to its own joint configuration without any explicit kinematic correspondence between the bodies.
  • The diffusion input during training carries most of the benefit: removing it raises average axis errors by roughly 31 mm, 77 mm, and 46 mm, so label-noise conditioning is the component that resolves redundant joint configurations.
  • RGB images are a richer condition for imitation than 2D or 3D pose skeletons, since replacing the image with pose estimates increases end-effector errors.
  • The 10 ms single-pass inference time on a consumer GPU makes the method usable for real-time imitation instead of offline batch processing.
  • Task generalisation is real but task-dependent: on the EMIL dataset, held-out tasks are imitated within 2.4 to 4.6 cm on average, with the closest tasks (lift and scoot) reaching around 1 cm.

Reading between the lines

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

  • If the reversed-imitation trick generalises, it turns dataset collection upside down: instead of manually aligning human and robot demonstrations, a lab can record a robot's scripted motions and ask humans to copy them, yielding inexpensive paired training data for many robot morphologies.
  • The same architecture should be testable on full-body imitation, since the two-arm result suggests the bottleneck is paired data and joint redundancy rather than the diffusion mechanism itself.
  • Because the model outputs joint values in one pass, it could serve as a fast inverse-kinematics prior or as a demonstration-conditioned reward generator in reinforcement learning, although the paper does not test either use.
  • A natural stress test would be moving the camera to a third-person viewpoint: the DIRI condition is recorded through the robot's own eye, so the claimed end-to-end perception-to-action coupling may depend on a fixed egocentric camera geometry.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents DIRIGENt, a diffusion-based U-Net that takes a single RGB image of a human demonstration and directly outputs 26 joint values for the NICOL robot. To obtain paired human--robot data, the authors collect the DIRI dataset by having ten human participants imitate slow, predictable robot arm movements; the robot's own joint trajectory is used as ground truth. The model is evaluated on a random 90:10 split of DIRI, on a held-out-participant fold (fold 10), on the EMIL robot-task dataset, and through a series of ablations that vary pose-estimation conditions, Cartesian versus joint losses, diffusion noise, iterative denoising, and temporal context. The authors report end-effector errors of 2/3/2 mm on the random split and 4.3/6.0/4.1 cm on the held-out participant fold, and claim that the full model outperforms the HoRoPose baseline.

Significance. The problem of learning robot joint configurations directly from RGB images of human demonstrations is relevant, and the paper contains several useful contributions: the reversed-imitation data-collection protocol is a creative way to generate paired data; the architecture combines a conditioning U-Net with a kinematic chain in the loss; and the ablations are extensive and mostly support the stated design choices. The held-out-participant evaluation and the comparison with HoRoPose are useful checks. However, as detailed in the major comments, the headline generalization claims are not yet supported by the experiments, and the state-of-the-art comparison rests on a single baseline. If the authors add the missing natural-motion evaluation and strengthen the comparison, the contribution could become solid for the robotics-imitation community.

major comments (4)
  1. [Section 3, DIRI dataset] The central claim that DIRIGENt enables a robot to imitate humans is not established for self-initiated natural human demonstrations. In the data-collection protocol, human participants imitate a slow, predictable robot, and the model is trained to invert that specific imitation behavior. There is no experiment in which a person performs an arbitrary, self-initiated movement and the robot is evaluated on matching it. The held-out participant fold still uses participants who are imitating the robot. Please add an evaluation with natural human demonstrations (or a third-party human-motion dataset) and/or substantially weaken the abstract and introduction claims.
  2. [Section 4, random 90:10 split] The headline 2/3/2 mm errors come from a random 90:10 split of frames from continuous recordings. Because temporally adjacent frames are highly correlated, a random frame-level split can place near-duplicate poses in both training and test sets, so these numbers largely reflect interpolation within a recording rather than generalization. Please report results with a split that respects recording boundaries or temporal gaps (for example, training on some recordings and testing on held-out recordings), and present that number as the generalization figure alongside the fold-10 result.
  3. [Table 1, HoRoPose comparison] The state-of-the-art claim is based on a single baseline, HoRoPose, trained with its default hyperparameters for 20 epochs. With only one baseline and no evidence that 20 epochs is appropriate for HoRoPose, the comparison supports 'competitive' rather than 'outperforms existing state-of-the-art approaches.' At minimum, report a learning curve or a tuned result for HoRoPose, and ideally add a second baseline that is also evaluated under comparable conditions.
  4. [Section 4, EMIL experiments] The EMIL experiments use robot demonstrations of the NICO robot, not human demonstrations, so they demonstrate generalization across robot tasks, not human-to-robot imitation. This is a useful auxiliary result, but the text should not be read as evidence for the central human-imitation claim. Please clarify this distinction in the presentation of the EMIL results.
minor comments (6)
  1. [Section 3.1, loss function] The loss function is presented without an equation number and uses symbols (DDO, DTI, DTI(eef), TI, and II) that are not all defined at the point of use; please introduce the notation carefully and number the equation.
  2. [Section 4, relative errors] The reported relative errors (0.004, 0.006, 0.007) do not match the stated axis ranges: 2 mm / 400 mm = 0.005 for x, 3 mm / 500 mm = 0.006 for y, and 2 mm / 300 mm = 0.0067 for z. Please recompute or clarify the base ranges used.
  3. [Table 1] The metric names and units should be defined explicitly: it is not clear whether 'Joint Loss' is MSE in radians or another unit, and whether 'Cartesian Loss' is MSE in meters or squared meters. The table also mixes configurations and datasets; a separate column for the experimental setting would improve readability.
  4. [Section 3.1, hyperparameters] The hyperparameter section states that training was performed on 'commercial GPUs' without giving the exact hardware or training time; please report the GPU model and the total training duration for reproducibility, and state whether random seeds were fixed.
  5. [Figure 2] The caption refers to the 'Direct Imitation Dataset' while the paper consistently uses the acronym DIRI; please unify the terminology.
  6. [Section 4, Figure 4] The statement that 'the model manages to adapt to this error' is vague; please explain more concretely how Figure 4 demonstrates adaptation to human imitation error.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; dataset-protocol concerns are external-validity issues, not construction-level circularity.

full rationale

The paper's derivation is a conventional supervised-learning pipeline. The DIRI collection protocol defines the input as an RGB image of a human imitating the robot and the label as the robot's original joint values; Section 3.1 makes this explicit by writing the target as DTI and the loss as ω1·MSE(DDO, DTI) + ω2·MSE(DDO(eef), DTI(eef)). The model is therefore trained to invert a forward process (robot pose -> human perception -> human imitation -> camera image), which is an inverse-mapping task, not a definitional identity. The reported superiority over HoRoPose is obtained by training HoRoPose from scratch on the same splits for the same number of epochs, so the comparison is not manufactured by a fitted parameter renamed as a prediction. The ablations (pose estimation, Cartesian vs joint loss, diffusion input, iterative denoising, temporal overlaying) each remove or modify one component and measure the effect on the same held-out protocol; none of the ablations builds the result into the objective by construction. Self-citations to the authors' prior work (robot platforms, CycleIK, EMIL, perspective-transfer architecture) are used as resources or architectural inspiration and are not invoked as a uniqueness theorem or as a substitute for the empirical comparison. The reversed-imitation protocol and the random 90:10 temporal split do raise external-validity questions about natural human demonstrations and frame-level leakage, but those concern whether the benchmark measures the intended real-world task; they do not make the reported output equal to the input by construction. No circular step can be exhibited, so the score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central contribution is an empirical method plus a dataset. The method depends on several hand-set hyperparameters (loss weights, noise levels, denoising steps, training budget) and on the dataset's reversed-imitation protocol. The key domain assumption is that humans imitating a slow robot produce image-target pairs that represent how a robot should imitate natural human demonstrations. No new theoretical entities are introduced; the forward kinematic model is taken from an existing library.

free parameters (4)
  • Loss weights omega1 and omega2 = not reported
    These weight the joint MSE and end-effector MSE in the combined loss; their values are chosen by hand and not disclosed, but they directly affect the optimization objective.
  • Number of diffusion noise levels = 1000
    Cosine noise schedule with 1000 levels; a design hyperparameter chosen without reported sensitivity analysis.
  • Inference denoising steps = 1 (default), 50 (ablation)
    The default uses a single-step prediction to achieve 10 ms inference; the ablation shows 50 steps give slightly different performance.
  • Training epochs and batch size = 20 epochs, batch size 6
    Training budget chosen for all models; no learning curve or convergence analysis is shown.
assumptions (5)
  • domain assumption The DIRI reversed-imitation dataset is representative of real human demonstrations for the task of robotic imitation.
    All training and evaluation data comes from humans deliberately imitating the robot's pre-recorded movements (Section 3). The central claim that DIRIGENt can imitate humans in general rests on this representativeness.
  • domain assumption Human error in imitating the robot is acceptable and can be learned as part of the mapping.
    The loss in Section 3.1 uses the robot's original pose as ground truth even though the human's imitation is the input image; the paper acknowledges this error in Section 4.
  • domain assumption The forward kinematic model for NICOL's arms is accurate enough for the Cartesian loss.
    The kinematic chain from PyTorch Kinematics is used to compute end-effector positions in the loss; any kinematic error propagates to the optimization (Section 3.1).
  • domain assumption MSE on joint values and end-effector positions is a valid proxy for imitation quality.
    No user study or task success rate is reported; all conclusions are based on these two MSE terms.
  • standard math Denoising diffusion models can be conditioned on auxiliary inputs and trained to predict the target directly.
    The method builds on Ho et al. (2020) and Nichol & Dhariwal (2021); this is accepted background in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIRIGENt: End-To-End Robotic Imitation of Human Demonstrations Based on a Diffusion Model." pith.science (2026). https://pith.science/paper/UKFBP2G4

@misc{pith2026250116800,
  author       = {Pith},
  title        = {Pith review of: DIRIGENt: End-To-End Robotic Imitation of Human Demonstrations Based on a Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UKFBP2G4}},
  note         = {Machine review of arXiv:2501.16800}
}
read the original abstract

There has been substantial progress in humanoid robots, with new skills continuously being taught, ranging from navigation to manipulation. While these abilities may seem impressive, the teaching methods often remain inefficient. To enhance the process of teaching robots, we propose leveraging a mechanism effectively used by humans: teaching by demonstrating. In this paper, we introduce DIRIGENt (DIrect Robotic Imitation GENeration model), a novel end-to-end diffusion approach that directly generates joint values from observing human demonstrations, enabling a robot to imitate these actions without any existing mapping between it and humans. We create a dataset in which humans imitate a robot and then use this collected data to train a diffusion model that enables a robot to imitate humans. The following three aspects are the core of our contribution. First is our novel dataset with natural pairs between human and robot poses, allowing our approach to imitate humans accurately despite the gap between their anatomies. Second, the diffusion input to our model alleviates the challenge of redundant joint configurations, limiting the search space. And finally, our end-to-end architecture from perception to action leads to an improved learning capability. Through our experimental analysis, we show that combining these three aspects allows DIRIGENt to outperform existing state-of-the-art approaches in the field of generating joint values from RGB images.

Figures

Figures reproduced from arXiv: 2501.16800 by the authors.

Figure 1
Figure 1. The robot NICOL is on the right side with a human demon [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of datasets: on the left is the EMIL dataset and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Our architecture. On the left side the differences between training and inference are shown as well as all inputs to the model, on the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The end effector position in the label, the model’s prediction [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: On the left is the human demonstration, and on the right is [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 30 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    Human to robot whole-body motion transfer

    Miguel Arduengo, Ana Arduengo, Adri \`a Colom \'e , Joan Lobo-Prat, and Carme Torras. Human to robot whole-body motion transfer. In Humanoids IEEE-RAS 20th International Conference on Humanoid Robots , pages 299--305. IEEE, 2021

    work page 2021

  3. [3]

    Human-to-robot imitation in the wild

    Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 , 2022

    arXiv 2022

  4. [4]

    Real-time holistic robot pose estimation with unknown states

    Shikun Ban, Juling Fan, Xiaoxuan Ma, Wentao Zhu, Yu Qiao, and Yizhou Wang. Real-time holistic robot pose estimation with unknown states. In European Conference on Computer Vision , pages 1--17. Springer, 2025

    work page 2025

  5. [5]

    Ilvr: Conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 14347--14356. IEEE, 2021

    work page 2021

  6. [6]

    Openmmlab pose estimation toolbox and benchmark

    MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020

    2020

  7. [7]

    Imitation of human motion on a humanoid robot using non-linear optimization

    Martin Do, Pedram Azad, Tamim Asfour, and Rudiger Dillmann. Imitation of human motion on a humanoid robot using non-linear optimization. In Humanoids 8th IEEE-RAS International Conference on Humanoid Robots , pages 545--552. IEEE, 2008

    work page 2008

  8. [8]

    Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot

    Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. In Towards Generalist Robots: Learning Paradigms for Scalable Skill Acquisition , 2023

    work page 2023

Show all 42 references
  1. [9]

    One-shot visual imitation learning via meta-learning

    Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on Robot Learning , pages 357--368. PMLR, 2017

    2017

  2. [10]

    Domain adaption as auxiliary task for sim-to-real transfer in vision-based neuro-robotic control

    Connor Gaede, Jan-Gerrit Habekost, and Stefan Wermter. Domain adaption as auxiliary task for sim-to-real transfer in vision-based neuro-robotic control. In International Joint Conference on Neural Networks, Yokohama, Japan. , Jul 2024

    2024

  3. [11]

    Cycleik: Neuro-inspired inverse kinematics

    Jan-Gerrit Habekost, Erik Strahl, Philipp Allgeuer, Matthias Kerzel, and Stefan Wermter. Cycleik: Neuro-inspired inverse kinematics. In International Conference on Artificial Neural Networks , pages 457--470. Springer, 2023

    2023

  4. [12]

    Inverse kinematics for neuro-robotic grasping with humanoid embodied agents

    Jan-Gerrit Habekost, Connor G \"a de, Philipp Allgeuer, and Stefan Wermter. Inverse kinematics for neuro-robotic grasping with humanoid embodied agents. arXiv preprint arXiv:2404.08825 , 2024

    2024 arXiv

  5. [13]

    Learning human-to-humanoid real-time whole-body teleoperation

    Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. arXiv preprint arXiv:2403.04436 , 2024

    2024 arXiv

  6. [14]

    Embodied multi-modal interaction in language learning: the emil data collection

    Stefan Heinrich, Matthias Kerzel, Erik Strahl, and Stefan Wermter. Embodied multi-modal interaction in language learning: the emil data collection. In ICDL-EpiRob Workshop on Active Vision, Attention, and Learning (ICDL-Epirob AVAL) , page 2p, 2018

    2018

  7. [15]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems , 33:6840--6851, 2020

    2020

  8. [16]

    Imitation learning: A survey of learning methods

    Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR) , 50(2):1--35, 2017

    2017

  9. [17]

    Nico—neuro-inspired companion: A developmental humanoid robot platform for multimodal interaction

    Matthias Kerzel, Erik Strahl, Sven Magg, Nicol \'a s Navarro-Guerrero, Stefan Heinrich, and Stefan Wermter. Nico—neuro-inspired companion: A developmental humanoid robot platform for multimodal interaction. In 26th IEEE International Symposium on Robot and Human Interactive Co...

    2017

  10. [18]

    Nicol: A neuro-inspired collaborative semi-humanoid rrobot that bridges social interaction and reliable manipulation

    Matthias Kerzel, Philipp Allgeuer, Erik Strahl, Nicolas Frick, Jan-Gerrit Habekost, Manfred Eppe, and Stefan Wermter. Nicol: A neuro-inspired collaborative semi-humanoid rrobot that bridges social interaction and reliable manipulation. IEEE access , 11:123531--123542, 2023

    2023

  11. [19]

    Real-time imitation of human whole-body motions by humanoids

    Jonas Koenemann, Felix Burget, and Maren Bennewitz. Real-time imitation of human whole-body motions by humanoids. In IEEE International Conference on Robotics and Automation (ICRA) , pages 2806--2812. IEEE, 2014

    2014

  12. [20]

    Imitation from observation: Learning to imitate behaviors from raw video via context translation

    YuXuan Liu, Abhishek Gupta, Pieter Abbeel, and Sergey Levine. Imitation from observation: Learning to imitate behaviors from raw video via context translation. In IEEE International Conference on Robotics and Automation (ICRA) , pages 1118--1125. IEEE, 2018

    2018

  13. [21]

    Diffusion action segmentation

    Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In IEEE/CVF International Conference on Computer Vision , pages 10139--10149, 2023

    2023

  14. [22]

    Rtmo: Towards high-performance one-stage real-time multi-person pose estimation

    Peng Lu, Tao Jiang, Yining Li, Xiangtai Li, Kai Chen, and Wenming Yang. Rtmo: Towards high-performance one-stage real-time multi-person pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1491--1500, 2024

    2024

  15. [23]

    Foundations for a new science of learning

    Andrew N Meltzoff, Patricia K Kuhl, Javier Movellan, and Terrence J Sejnowski. Foundations for a new science of learning. Science , 325(5938):284--288, 2009

    2009

  16. [24]

    Nehaniv and Kerstin Dautenhahn

    Chrystopher L. Nehaniv and Kerstin Dautenhahn. The Correspondence Problem . The MIT Press, 06 2002

    2002

  17. [25]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning , pages 8162--8171. PMLR, 2021

    2021

  18. [26]

    A real-time human imitation system using kinect

    Yongsheng Ou, Jianbing Hu, Zhiyang Wang, Yiqun Fu, Xinyu Wu, and Xiaoyun Li. A real-time human imitation system using kinect. International Journal of Social Robotics , 7:587--600, 2015

    2015

  19. [27]

    Quaternet: A quaternion-based recurrent model for human motion

    Dario Pavllo, David Grangier, and Michael Auli. Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 , 2018

    2018 arXiv

  20. [28]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-assisted Intervention--MICCAI: 18th International Conference, Munich, Germany , pages 234--241. Springer, 2015

    2015

  21. [29]

    Time-contrastive networks: Self-supervised learning from video

    Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In IEEE International Conference on Robotics and Automation (ICRA) , pages 1134--1141. IEEE, 2018

    2018

  22. [30]

    Multiple interactions made easy (mime): Large scale demonstrations data for imitation

    Pratyusha Sharma, Lekha Mohan, Lerrel Pinto, and Abhinav Gupta. Multiple interactions made easy (mime): Large scale demonstrations data for imitation. In Conference on Robot Learning , pages 906--915. PMLR, 2018

    2018

  23. [31]

    Third-person visual imitation learning via decoupled hierarchical controller

    Pratyusha Sharma, Deepak Pathak, and Abhinav Gupta. Third-person visual imitation learning via decoupled hierarchical controller. Advances in Neural Information Processing Systems , 32, 2019

    2019

  24. [32]

    Avid: Learning multi-stage tasks via pixel-level translation of human videos

    Laura Smith, Nikita Dhawan, Marvin Zhang, Pieter Abbeel, and Sergey Levine. Avid: Learning multi-stage tasks via pixel-level translation of human videos. arXiv preprint arXiv:1912.04443 , 2019

    1912 arXiv

  25. [33]

    Diffusing in someone else's shoes: Robotic perspective taking with diffusion

    Josua Spisak, Matthias Kerzel, and Stefan Wermter. Diffusing in someone else's shoes: Robotic perspective taking with diffusion. arXiv preprint arXiv:2404.07735 , 2024

    2024 arXiv

  26. [34]

    Robotic imitation of human actions

    Josua Spisak, Matthias Kerzel, and Stefan Wermter. Robotic imitation of human actions. arXiv preprint arXiv:2401.08381 , 2024

    2024 arXiv

  27. [35]

    Teleoperation of a humanoid robot using full-body motion capture, example movements, and machine learning

    Christopher Stanton, Anton Bogdanovych, and Edward Ratanasena. Teleoperation of a humanoid robot using full-body motion capture, example movements, and machine learning. In Australasian Conference on Robotics and Automation , volume 8, page 51, 2012

    2012

  28. [36]

    On human motion imitation by humanoid robot

    Wael Suleiman, Eiichi Yoshida, Fumio Kanehiro, Jean-Paul Laumond, and Andr \'e Monin. On human motion imitation by humanoid robot. In IEEE International Conference on Robotics and Automation (ICRA) , pages 2697--2704. IEEE, 2008

    2008

  29. [37]

    Integrating reinforcement learning with human demonstrations of varying ability

    Matthew E Taylor, Halit Bener Suay, and Sonia Chernova. Integrating reinforcement learning with human demonstrations of varying ability. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 2 , pages 617--624, 2011

    2011

  30. [38]

    One-shot imitation from observing humans via domain-adaptive meta-learning

    Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. arXiv preprint arXiv:1802.01557 , 2018

    2018 arXiv

  31. [39]

    Imitation system of humanoid robots and its applications

    Ze-Feng Zhan and Han-Pang Huang. Imitation system of humanoid robots and its applications. IEEE Open Journal of Circuits and Systems , 4:15--24, 2022

    2022

  32. [40]

    A real-time upper-body robot imitation system

    Zhijun Zhang, Yaru Niu, Lingdong Kong, Shuyang Lin, and Hao Wang. A real-time upper-body robot imitation system. International Journal of Robotics and Control , 2:49--56, 2019

    2019

  33. [41]

    PyTorch Kinematics , February 2024

    Sheng Zhong, Thomas Power, Ashwin Gupta, and Peter Mitrano. PyTorch Kinematics , February 2024

    2024

  34. [42]

    Motionbert: A unified perspective on learning human motion representations

    Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In IEEE/CVF International Conference on Computer Vision , pages 15085--15099, 2023

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.