HumiMimic: Mimetic Embodiment Transfer of Human-Humanoid Interaction from In-the-Wild Videos
Authors: Nghia Le, Khang Nguyen, Ba Long Tran, Thi Quynh Nhu Pham, Binh Nguyen, Tuong Do, Erman Tjiputra, Hoan Nguyen, Tung D. Ta, Baoru Huang, Hai-Nguyen Nguyen, Anh Nguyen
Organizations: Hann · AIOZ Ltd., Singapore · MBZUAI, Abu Dhabi, United Arab Emirates · RMIT University, Vietnam Campus · University of Liverpool, UK · National Tsing Hua University, Taiwan · University of Information Technology, VNUHCM, Vietnam. · Keio University, Mita, Minato-ku, Tokyo, Japan
Enabling a humanoid to interact naturally with human partners requires preserving not only individual motions but also the spatial, semantic, and contact relationships underlying the interaction (i.e., hugging, dancing, hand-shaking, etc.). Furthermore, a practical solution should generalize across various humanoid embodiments with distinct morphologies and kinematic constraints. To tackle this challenge, we propose HumiTransfer, a mimetic embodiment transfer framework that converts in-the-wild human-human interaction videos into physically plausible human-humanoid interactions. HumiTransfer first reconstructs 3D human motion with consistent foot-ground contact, then transfers one participant to a target humanoid embodiment via interaction modeling, using weighted joint-relative directions and anatomical offsets to prioritize interaction-critical inter-body relationships, while post-retargeting optimization enforces balance, contact, and collision avoidance. Subsequently, we apply HumiTransfer to 1,200 in-the-wild videos, producing approximately 500,000 frames of diverse interactions for the 26-DoF Unitree R1 Edu and the 23-DoF HighTorque Mini-hi humanoid robot, resulting in our HumiMimic dataset. Across both robots, HumiTransfer achieves superior motion fidelity, preserves inter-body relationships, and maintains contact accuracy compared with existing baselines. Moreover, we demonstrate that HumiMimic supports motion generation methods for synthesizing interactive humanoid motions following a human partner. Finally, our real robot deployments confirm the physical executability of the transferred motions in real-world settings. Our project is available at https://aioznetwork.github.io/HumiMimic/
Figures & tables
TABLE I: Key Attributes of Interaction Datasets: Comparisons of key attributes between HumiMimic dataset and existing human-motion and humanoid-motion datasets in terms of collection setting, scale, action diversity, and supported humanoid embodiments.
Fig. 2: Overview of HumiTransfer: (A) Given an in-the-wild video of a human-human interaction, HumiTransfer tracks human participants, detects body keypoints, reconstructs mesh-based full-body frames, and refines the resulting frames to ensure consistent ground-foot contact (Sec. III-A ). (B) Next, the poses of one participant are initialized on the target humanoid through constrained inverse kinematics. HumiTransfer then represents the interaction semantics using joint-relative directions JtR→P and anatomical offsets AtR→P . Within this process, semantic and distance-based weights, WsemH→P and Wdis,tH→P , are used to prioritize the inter-body relationships that are important to the observed interaction, allowing their alignment across human and humanoid embodiments (Sec. III-B ). (C) The retargeted pose is further refined to maintain humanoid balance and suppress unintended inter-body collisions while preserving intended contacts and interaction dynamics, yielding the final human-humanoid interaction (Sec. III-C ). By applying HumiTransfer onto the Unitree R1 Edu and the HighTorque Mini-hi, we obtain HumiMimic , our human-humanoid interaction dataset.
Fig. 3: Dataset Statistics: Distribution of interaction lengths (in frames) and interaction diversity (in percentages) from the collected in-the-wild human-human interaction videos.
Fig. 4: Qualitative Comparisons of Retargeting Results: Given the human-human interaction to retarget to different humanoids, naive IK [ 25 ] produces suboptimal leg and head poses on the Unitree R1 Edu and foot-ground penetration on the HighTorque Mini-hi. Meanwhile, GMR [ 23 ] and HumiTransfer - (Ours) preserve robot poses but lose contact with the human partner, whereas PAIR [ 37 ] introduces distorted poses on both humanoids. HumiTransfer + (Ours) preserves embodiment-specific poses and maintains contact with the human partner.
TABLE II: Effects of Foot-Ground Consistency Fine-tuning: Quantitative comparisons between SMPL-based fitting and itself with our proposed foot-ground consistency fine-tuning (Eq. ( 3 )) in terms of Efoot , Eground , and SSM on HumiMimic .
Method
Unitree R1 Edu
HighTorque Mini-hi
MPJPE ( ↓ )
IRD ( ↓ )
CF1 ( ↑ )
MPJPE ( ↓ )
IRD ( ↓ )
CF1 ( ↑ )
Naive IK [ 25 ]
37.85
16.82
13.84
41.31
20.67
11.42
GMR [ 23 ]
11.91
5.73
69.12
10.49
5.30
61.97
PAIR [ 37 ]
9.37
4.16
74.54
12.79
6.24
59.41
HumiTransfer -
10.13
4.08
80.90
11.63
6.14
65.29
HumiTransfer +
6.31
2.84
81.97
7.58
3.67
78.15
TABLE III: Retargeting Performance on Different Humanoids: Quantitative evaluations in terms of MPJPE, IRD, and CF 1 on HumiMimic between our methods against naive IK [ 25 ] , GMR [ 23 ] , and PAIR [ 37 ] on the Unitree R1 Edu and the HighTorque Mini-hi. HumiTransfer - and HumiTransfer + denote for without and with the weighting matrices (Eq. 6 and Eq. 7 ).
Fig. 5: Qualitative Samples of Retargeted Human-Humanoid Interactions: By leveraging internet-scale human-human interaction videos, HumiMimic exhibits numerous human-humanoid interactions. More samples are available in our supplementary video.
Method
Unitree R1 Edu
HighTorque Mini-hi
Diversity ( ↑ )
Diversity ( ↑ )
Human3.6M [ 5 ]
3.8251
3.7963
AMASS [ 4 ]
5.4834
5.4859
3DPW [ 7 ]
4.0173
3.9130
HumiMimic
6.2427
6.2285
TABLE IV: Motion Diversity of Datasets on Different Humanoids: Diversity scores [ 54 ] of human-humanoid interactions on HumiMimic compared to Human3.6M [ 5 ] , AMASS [ 4 ] , and 3DPW [ 7 ] on the Unitree R1 Edu and the HighTorque Mini-hi.
Fig. 6: Qualitative Samples of Generative Human-Humanoid Interaction: Qualitative comparison of humanoid-human interaction generation on the test set of HumiMimic . The context of the input is “ A man is kicking a ball ”; the humanoid is expected to move to receive the ball as it is passed. Duolando [ 59 ] achieves a more natural and reactive response to follower motions, while InterGen [ 58 ] either mimics the partner’s motion or responds with an unsuitable motion. More qualitative results of other baselines are provided in our supplementary video.
Method
Unitree R1 Edu
HighTorque Mini-hi
FID k ( ↓ )
FID cd ( ↓ )
fcontact ( ↑ )
FID k ( ↓ )
FID cd ( ↓ )
fcontact ( ↑ )
ComMDM [ 57 ]
36.07
15.93
31.09%
38.33
18.30
29.07%
InterFormer [ 56 ]
30.16
14.66
42.97%
28.64
13.78
37.49%
InterGen [ 58 ]
34.62
17.72
48.91%
30.98
10.86
39.37%
ReMos [ 55 ]
25.49
11.66
46.79%
31.51
12.94
47.56%
Duolando [ 59 ]
24.96
9.07
51.75%
27.24
11.53
46.77%
TABLE V: Generating Humanoid Motion from Human Motion Results. Quantitative results on humanoid-human interaction between ComMDM [ 57 ] , InterFormer [ 56 ] , InterGen [ 58 ] , ReMos [ 55 ] , and Duolando [ 59 ] . All experiments run on the test set of our HumiMimic dataset.
Fig. 7: Visualization of CoM Trajectories on Different Humanoid Embodiments: Qualitative comparison of the original human motion (top) and the retargeted motions to Unitree R1 Edu (middle) and HighTorque Mini-hi (bottom) during the kicking interaction. The yellow markers visualize CoM trajectories of humans and humanoids, showing their geometric similarities to the human’s CoM trajectories.
Fig. 8: Real-Robot Deployment on the Unitree R1 Edu: Two deployed human-humanoid interaction sequences on the Unitree R1 Edu: (a) hugging and (b) hand-shaking with overlaid snapshots depict the temporal evolution of the interacting motion on both the humanoid and its human partner. The interacting sequences are safe, and the humanoid does not cause harm in either scenario.
Cross-embodiment video generation aims to transfer motions across different humanoid embodiments, such as human-to-robot and robot-to-robot, enabling scalable data generation for embodied intelligence. A major challenge in this setting is that motion dynamics are partly transferable across embodiments, whereas appearance and morphology remain embodiment-specific. Existing approaches often entangle these factors, and many require paired data for every target embodiment, which limits scalability to new robots. We present OmniHumanoid, a framework that factorizes transferable motion learning and embodiment-specific adaptation. Our method learns a shared motion transfer model from motion-aligned paired videos spanning multiple embodiments, while adapting to a new embodiment using only unpaired videos through lightweight embodiment-specific adapters. To reduce interference between motion transfer and embodiment adaptation, we further introduce a branch-isolated attention design that separates motion conditioning from embodiment-specific modulation. In addition, we construct a synthetic cross-embodiment dataset with motion-aligned paired videos rendered across diverse humanoid assets, scenes, and viewpoints. Experiments on both synthetic and real-world benchmarks show that OmniHumanoid achieves strong motion fidelity and embodiment consistency, while enabling scalable adaptation to unseen humanoid embodiments without retraining the shared motion model.
Captured human-object interactions provide rich supervision for humanoid loco-manipulation, but they are sparse, heterogeneous, and not directly executable by robots. We introduce InterMimicGen, a self-evolving motion-imitation framework in which robot motion data and a tracking policy improve each other. First, we consolidate motion-captured human-object interaction datasets and retarget them into humanoid robot references while preserving whole-body coordination and dexterous hand-object relationships. This produces a large and diverse humanoid robot reference collection for dexterous whole-body loco-manipulation. Second, we train a physics-based generalist tracker that executes these references in simulation on a humanoid with dexterous hands, covering a scale and diversity beyond prior humanoid tracking systems for loco-manipulation. Third, we close a data flywheel: each round makes small, task-preserving changes to where an interaction takes place and how the body performs it, fine-tunes the tracker on them, and keeps only the variants whose simulated execution completes the task, which seed the next round. With more iterations, these small edits compound into broader coverage around the sparse original demonstrations while preserving task semantics and motion quality. Experiments show contact-preserving retargeting across robot configurations, broad tracking with a single generalist policy, executable motions that keep growing over augmentation rounds, and transfer to real robots. InterMimicGen provides a unified path from heterogeneous human demonstrations to a continually expanding motion resource for humanoid robot learning.
Learning from demonstration (LfD) has enabled humanoid robots to acquire diverse whole-body skills, but extending this paradigm to human-object interaction (HOI) is limited by the availability of robot-compatible interaction references. We present HOI-Retarget, a contact-centric retargeting method that transfers HOI onto a humanoid robot for large-scale motion-data generation. Its windowed trajectory optimization uses every labeled contact as a target in the object frame, balancing body tracking, foot support and smoothness under the robot's kinematic limits. The method can augment a single demonstration across object sizes, absorb contacts reconstructed from monocular video, and extend to several robots manipulating one object. We publicly release the code and the retargeted motion dataset.
Jihwan Shin, Adrià López Escoriza, Junzhe He +2
Robotic Systems Lab, ETH Zürich, 8092 Zürich, Switzerland