Literature
Display
← All papers

PHUMA: Physically Reliable Humanoid Locomotion Dataset

Kyungmin Lee · Sibeen Kim · Youngdo Lee · Minho Park · Hyunseung Kim · Dongyoon Hwang · Donghu Kim · Hojoon Lee · Jaegul Choo

Preprint posted · CoRL 2026 (accepted) ↗

Robotics15 min text · CondensedOriginal paper ↗
Contents

Overview

Complete original PDF figure S0.F1.
Figure 1: Physical Reliability of Humanoid-X vs. PHUMA. Each column illustrates four failure modes: joint violation, floating, penetration, and skating. Humanoid-X [1] (top row) often exhibits these issues due to direct video-to-motion conversion, while PHUMA (bottom row) mitigates those violations through careful data curation and physically reliable retargeting.

At a glance

PHUMA combines physics-aware filtering of human motion with PhySINK, a shape-adaptive retargeting optimizer that penalizes joint-limit violations, poor foot grounding and skating. The resulting dataset contains about 76,000 clips and 73 hours. Policies trained on it improve aggregate tracking results on G1 and H1-2 in simulation, with a smaller hardware study on G1. These are kinematic quality checks and empirical tracking results, rather than a guarantee that every motion satisfies full robot dynamics.

1 Introduction

Motion imitation depends on both the coverage of the human demonstrations and their suitability for a different body. Motion capture is relatively clean but limited; internet-video reconstruction scales collection while introducing jitter, floating and penetration. PHUMA addresses both stages: reject unsuitable human trajectories, then optimize their transfer to the robot.

This edition follows the 35-page arXiv v3 dated 12 October 2026, including its expanded ablations. The catalog retains the original 2025 posting date. The reported 73 hours are roughly 3.5 times the 20.9 hours deemed feasible in AMASS; this comparison concerns the paper’s selected feasible portions, not the entirety of every original dataset.

Complete original PDF figure S1.F2.
Figure 2: Overview of datasets and performance. (a) Composition of feasible and infeasible human motion sources in each dataset. (b) Physical reliability of each dataset; AMASS is retargeted using a standard learning-based inverse kinematics method. (c) Success rate on unseen motions. (d) Motion tracking error on real Unitree G1.

2 Related Work

The work links human-motion datasets and robot retargeting. Its claim is that improving both the source trajectories and their morphology-dependent mapping gives more useful imitation targets.

2.1 Human Motion Data

AMASS and LaFAN1 supply motion-capture data; Humanoid-X illustrates video-driven scale. PHUMA filters motion directly using contact and balance-related heuristics, without requiring a pretrained tracking policy to decide which clips survive. This avoids one source of policy-dependent selection, but its hand-set flat-ground criteria can still exclude feasible behaviors outside the intended scope.

2.2 Humanoid Motion Retargeting

Rigid scaling and standard inverse kinematics can distort human-to-robot proportions. Shape adaptation improves matching but does not itself impose contact or joint constraints. GMR uses heuristic scaling and IK, while dynamics-aware approaches such as SPIDER add optimization beyond kinematic retargeting. PhySINK supplies a contact-aware kinematic initialization; it does not replace the latter dynamics optimization.

3 Method

The pipeline collects human motion, smooths and filters it, retargets it with physical penalties, and trains tracking policies. LaFAN1, LocoMuJoCo and collected video supplement the Humanoid-X sources. Figure 3 preserves the complete four-stage diagram and loss illustrations.

Complete original PDF figure S3.F3.
Figure 3: Overview of the PHUMA pipeline. Our four-stage pipeline for motion imitation learning includes: (1) Motion Curation, where we filter out problematic motions from a diverse dataset; (2) Motion Retargeting, where the filtered motions are retargeted to the humanoid using PhySINK; (3) Policy Learning, where a policy is trained to imitate the retargeted motions; and (4) Inference, where the trained policy is used to control the humanoid, enabling it to imitate motions from unseen videos processed by a video-to-motion model.

3.1 Physics-Aware Motion Curation

Curation. Smooth high-frequency noise, estimate one ground plane by majority voting over foot-vertex heights, and shift the whole trajectory to that plane. Four-second clips are checked for excessive root jerk, inadequate foot contact, implausible pelvis height and pelvis/spine displacement from the base of support; failing sub-clips are discarded.

These criteria target flat-ground locomotion. Pelvis and spine joints approximate the center of mass, and a horizontal foot/ankle hull approximates support. Such proxies do not constitute a dynamic balance test. Their exclusions deliberately include sitting and other scene-dependent motions as well as reconstruction errors.

3.2 Physics-Aware Motion Retargeting

Fidelity. Optimize robot joint positions and root translation against a robot-shaped SMPL-X target. The objective combines global position matching, local adjacent-joint relationships and temporal smoothness. The printed local “orientation” term uses displacement-vector inner products without an explicit normalization, so it should not be silently reinterpreted as a normalized angular error.

Physical penalties. Penalize joint positions and velocities beyond 98% of mechanical limits. Contact ratios weight squared foot-height error and horizontal foot speed. These soft losses reduce artifacts without guaranteeing exact contact or dynamic feasibility.

Evaluation. Report the fraction of frames satisfying fidelity (position under 10 cm and link orientation under 10°), joint limits, contact height within 1 cm above/below ground, or contact speed under 10 cm/s. On G1, PhySINK reaches 94.8% fidelity, 100% joint feasibility, 99.9% non-floating, 96.8% non-penetration and 89.7% non-skating. Thus residual penetration/skating remain. GMR’s 100% non-penetration accompanies only 14.7% non-floating.

Equations 1–8
ℒglobal-match=∑t∑i‖piSMPL-X​(t)−piHumanoid​(t)‖1\mathcal{L}_{\text{global-match}}=\sum_{t}\sum_{i}\left\|p_{i}^{\text{SMPL-X}}(t)-p_{i}^{\text{Humanoid}}(t)\right\|_{1} (1)
ℒlocal-match\displaystyle\mathcal{L}_{\text{local-match}} =∑t∑i≠jmi​j​‖Δ​pi​jSMPL-X​(t)−Δ​pi​jHumanoid​(t)‖22⏟position\displaystyle=\sum_{t}\sum_{i\neq j}m_{ij}\underbrace{\left\|\Delta p_{ij}^{\text{SMPL-X}}(t)-\Delta p_{ij}^{\text{Humanoid}}(t)\right\|_{2}^{2}}_{\text{position}} (2)
+∑t∑i≠jmi​j(1−⟨Δ​pi​jSMPL-X​(t),Δ​pi​jHumanoid​(t)⟩)⏟orientation\displaystyle+\sum_{t}\sum_{i\neq j}m_{ij}\underbrace{\left(1-\left\langle\Delta p_{ij}^{\text{SMPL-X}}(t),\ \Delta p_{ij}^{\text{Humanoid}}(t)\right\rangle\right)}_{\text{orientation}}
ℒsmooth=∑t∥q˙t−2q˙t+1+q˙t+2∥1+∑t∥γ˙t−2γ˙t+1+γ˙t+2∥1\mathcal{L}_{\text{smooth}}=\sum_{t}\bigl\lVert\dot{q}_{t}-2\dot{q}_{t+1}+\dot{q}_{t+2}\bigr\rVert_{1}+\sum_{t}\bigl\lVert\dot{\gamma}_{t}-2\dot{\gamma}_{t+1}+\dot{\gamma}_{t+2}\bigr\rVert_{1} (3)
ℒFidelity=wglobal-match​ℒglobal-match+wlocal-match​ℒlocal-match+wsmooth​ℒsmooth\mathcal{L}_{\text{Fidelity}}=w_{\text{global-match}}\mathcal{L}_{\text{global-match}}+w_{\text{local-match}}\mathcal{L}_{\text{local-match}}+w_{\text{smooth}}\mathcal{L}_{\text{smooth}} (4)
ℒpos-violation\displaystyle\mathcal{L}_{\text{pos-violation}} =∑t[max⁡(0,qt−0.98​qmax)+max⁡(0,0.98​qmin−qt)]\displaystyle=\sum_{t}\left[\max(0,q_{t}-0.98q_{\text{max}})+\max(0,0.98q_{\text{min}}-q_{t})\right] (5)
ℒvel-violation\displaystyle\mathcal{L}_{\text{vel-violation}} =∑t[max⁡(0,q˙t−0.98​q˙max)+max⁡(0,0.98​q˙min−q˙t)]\displaystyle=\sum_{t}\left[\max(0,\dot{q}_{t}-0.98\dot{q}_{\text{max}})+\max(0,0.98\dot{q}_{\text{min}}-\dot{q}_{t})\right] (6)
ℒFeasibility\displaystyle\mathcal{L}_{\text{Feasibility}} =ℒpos-violation+ℒvel-violation\displaystyle=\mathcal{L}_{\text{pos-violation}}+\mathcal{L}_{\text{vel-violation}} (7)
ℒGround=∑i∈{LH,LT,RH,RT}∑tcti​‖pti​(z)‖22\mathcal{L}_{\text{Ground}}=\sum_{i\in\{\text{LH},\text{LT},\text{RH},\text{RT}\}}\sum_{t}c_{t}^{i}\left\|p_{t}^{i}(z)\right\|_{2}^{2} (8)
Table 1: Quantitative comparison and ablation study of retargeting methods. We evaluate performance on the G1 humanoid, showing the progressive impact of adding each of our proposed physical constraint losses. (H1-2 results in Appendix G)
Motion Fidelity (%) Joint Feasibility (%) Non-Floating (%) Non-Penetration (%) Non-Skating (%)
IK 27.6 91.7 55.6 47.8 59.7
GMR 56.3 81.8 14.7 100.0 67.7
SINK 94.8 95.9 96.4 14.9 55.4
+ Joint Feasibility Loss 94.9 100.0 96.4 14.8 55.6
+ Grounding Loss 94.9 100.0 99.9 97.2 53.6
+ Skating Loss = PhySINK 94.8 100.0 99.9 96.8 89.7
Equations 9–10
ℒSkate=∑i∈{LH,LT,RH,RT}∑tcti​‖p˙ti​(x,y)‖2\mathcal{L}_{\text{Skate}}=\sum_{i\in\{\text{LH},\text{LT},\text{RH},\text{RT}\}}\sum_{t}c_{t}^{i}\left\|\dot{p}_{t}^{i}(x,y)\right\|_{2} (9)
ℒPhySINK=ℒFidelity+wFeasibility​ℒFeasibility+wGround​ℒGround+wSkate​ℒSkate\mathcal{L}_{\text{PhySINK}}=\mathcal{L}_{\text{Fidelity}}+w_{\text{Feasibility}}\mathcal{L}_{\text{Feasibility}}+w_{\text{Ground}}\mathcal{L}_{\text{Ground}}+w_{\text{Skate}}\mathcal{L}_{\text{Skate}} (10)

4 Experiments

Experiments separately examine filtering, retargeting, complete datasets and hardware transfer. Their controls and evaluation thresholds differ, so each comparison answers a narrower question than a single universal benchmark ranking.

4.1 Experiment Setup

Train MaskedMimic tracking policies with PPO for simulation, and a modified BeyondMimic policy for hardware. G1 has 29 DoF; H1-2 uses 21, excluding wrists. Hold out about 7,500 PHUMA clips and evaluate another 504 collected videos reconstructed and retargeted into references.

Main simulation comparisons use a strict 0.15 m whole-body success threshold. Several later appendices use the conventional 0.5 m threshold and GMR references instead. Hardware lacks global-translation motion capture and therefore measures local MPJPE and joint velocity/acceleration errors; those results do not establish global path accuracy.

4.2 Effect of Physics-Aware Curation

To isolate curation, retarget every Humanoid-X variant with SINK and exclude pre-retargeted LaFAN1/LocoMuJoCo. Any individual filter improves tracking, while all filters reduce roughly 231 hours to 62 yet raise success from 50.6% to 87.0% on PHUMA Test and 39.1% to 79.6% on Unseen Video. This supports useful selection under the tested protocol, not an unconditional rule that less data is better.

Table 2: Imitation Performance: Effect of Physics-Aware Curation. MaskedMimic policies trained on Humanoid-X subsets with varying curation filters, on Unitree G1. All variants use SINK retargeting to isolate filtering effect. Hours: remaining dataset size after filtering.
Curation Filter PHUMA Test Unseen Video
Variant Hours Jerk FC Height BoS Total Stat. Ang. Vert. Horiz. Total Stat. Ang. Vert. Horiz.
Humanoid-X 231 ✗ ✗ ✗ ✗ 50.6 78.4 43.0 26.0 31.8 39.1 78.0 39.6 23.0 6.5
+ Jerk only 141 ✓ ✗ ✗ ✗ 83.0 91.5 77.6 72.3 88.5 77.6 100.0 78.8 62.8 61.5
+ FC only 123 ✗ ✓ ✗ ✗ 82.9 91.3 77.8 70.4 89.1 78.2 99.1 79.3 61.7 65.9
+ Height only 136 ✗ ✗ ✓ ✗ 80.6 90.7 74.6 69.3 85.1 74.0 98.3 76.4 61.7 50.5
+ BoS only 110 ✗ ✗ ✗ ✓ 81.3 90.2 76.2 72.0 84.4 74.8 96.6 79.8 54.3 57.1
All filters 62 ✓ ✓ ✓ ✓ 87.0 93.0 82.7 81.8 90.3 79.6 98.3 82.8 59.6 69.2

4.3 PhySINK Retargeting Method Effectiveness

Retarget identical AMASS motions using IK, GMR, SINK or PhySINK and train separate policies. PhySINK yields 79.5%/72.8% aggregate success on PHUMA Test/Unseen Video, versus SINK’s 76.2%/70.2%. The gains are strongest in some angular and vertical categories; the appendix checks that the advantage also appears with GMR-generated evaluation references.

Table 3: Motion tracking performance across retargeting approaches. Success rates of policies trained on AMASS data retargeted by IK, GMR, SINK, and PhySINK, evaluated on the G1 humanoid robot using two test sets: PHUMA Test and Unseen Video. (H1-2 results in Appendix G)
PHUMA Test Unseen Video
Retarget Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
IK 52.8 75.3 43.9 24.3 44.2 54.6 89.3 54.6 32.7 43.3
GMR 61.7 80.2 51.7 32.5 74.0 59.3 87.1 64.5 31.9 40.7
SINK 76.2 88.5 72.1 56.8 66.8 70.2 90.7 75.0 62.7 44.1
PhySINK 79.5 89.9 76.1 61.1 69.5 72.8 93.3 78.2 65.6 47.3

4.4 PHUMA Dataset Effectiveness

PHUMA gives 92.7%/82.9% G1 success on the two test sets. This dataset comparison uses supplied LaFAN1/Humanoid-X targets and SINK-retargeted AMASS, so data volume, coverage and retargeting quality all differ. Appendix D.4 also compares AMASS retargeted with PhySINK, supporting an additional data benefit.

Table 4’s AMASS Unseen Video total is 78.2%, but Table 3 and Appendix Table 21 give 70.2% with the same category entries. The original values are preserved. PHUMA leads every G1 category in this table, but the broader claim of leading every category on both robots is contradicted by H1-2 horizontal results in Table 39.

Table 4: Motion tracking performance across datasets. Success rates of policies trained on LaFAN1, AMASS, Humanoid-X, and PHUMA, evaluated across motion categories on the G1 humanoid robot using two test sets: PHUMA Test and Unseen Video. (H1-2 results in Appendix G)
PHUMA Test Unseen Video
Dataset Hours Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
LaFAN1 2.4 46.1 66.1 36.2 24.0 42.5 28.4 46.9 28.4 19.6 10.5
AMASS 20.9 76.2 88.5 72.1 56.8 66.8 78.2 90.7 75.0 62.7 44.1
Humanoid-X 231.4 50.6 78.4 43.0 26.0 31.8 39.1 78.0 39.6 23.0 6.5
PHUMA 73.0 92.7 95.6 91.7 86.0 85.6 82.9 96.7 88.0 71.8 67.1

4.5 Real-World Experiment

Extend BeyondMimic from single-motion tracking with an encoder of future reference frames. Train in IsaacSim and deploy zero-shot to G1 on six held-out motions from each of four categories. PHUMA reduces aggregate local MPJPE from 41.0 to 34.3 mm, a 16.3% relative reduction, and reduces aggregate velocity/acceleration errors from 3.20/1.50 to 2.69/1.29.

The units for the latter errors are degrees/frame and degrees/frame². Improvement is not universal: horizontal acceleration error increases from 1.80 to 1.96. Only AMASS is tested as a hardware baseline; this 24-motion comparison does not measure global translation or establish performance on scene interactions.

Table 5: Real-World Tracking Results. Tracking errors of policies trained on AMASS and PHUMA, evaluated on 6 motions per category from the PHUMA test set, on the Unitree G1.
EmpjpeE_{\mathrm{mpjpe}} (mm) ↓\downarrow Edof​_​velE_{\mathrm{dof\_vel}} (deg/frame) ↓\downarrow Edof​_​accE_{\mathrm{dof\_acc}} (deg/frame2) ↓\downarrow
Dataset Total Stat. Ang. Vert. Hor. Total Stat. Ang. Vert. Hor. Total Stat. Ang. Vert. Hor.
AMASS 41.0 27.3 47.5 49.9 39.4 3.20 1.81 4.10 3.14 3.77 1.50 0.77 2.09 1.33 1.80
PHUMA 34.3 23.2 40.4 42.6 31.1 2.69 1.41 3.07 2.68 3.60 1.29 0.58 1.43 1.18 1.96

5 Limitations

The dataset focuses on flat terrain and excludes object interaction, sitting, climbing and uneven surfaces. Many video motions are discarded, constraining behavior coverage; the retained targets still contain artifacts. Hardware comparison is limited to AMASS because the Humanoid-X policy falls more often in simulation. Reliable kinematic references therefore remain one component of robust real-world control.

Complete original PDF figure S5.F4.
Figure 4: Real-World Deployment. Each row shows the reference motion (top), the motion tracked by the AMASS-trained policy (middle), and the motion tracked by the PHUMA-trained policy (bottom), demonstrating that PHUMA achieves closer alignment with the reference across all sequences.

6 Conclusion

The evidence favors jointly curating source motion and improving retargeting for this locomotion setting. Simulation covers two humanoids and multiple policy/evaluation variants; hardware confirms aggregate local tracking gains on one robot and a small selected motion set. Extending the approach to environmental contacts and full dynamics remains open.

Acknowledgments

The authors acknowledge Korean government, research-institute and GPU-support funding. Bibliographic and attribution material is retained.

Appendix

The appendix supplies curation parameters, dataset composition, retargeting comparisons, policy configurations, evaluation variants, ablations and H1-2 results.

Appendix A PHUMA Dataset

Dataset construction depends on explicit filtering choices, rather than treating all reconstructed human motion as directly executable robot motion.

A.1 Data Preprocessing

Preprocessing smooths trajectories, derives contact signals from the human mesh, estimates a shared ground level and removes unsuitable sub-clips before retargeting.

A.1.1 Low-Pass Noise Filtering for Motion Data

Resample to 30 Hz and apply a zero-phase fourth-order Butterworth low-pass filter. Cutoffs are 3 Hz for root translation and 6 Hz for global orientation/body pose. This is offline trajectory preprocessing, not a causal real-time filter.

A.1.2 Extracting Ground Contact Information

Select the 22 lowest SMPL-X vertices in each heel/toe region, 88 total; Table 6 retains every index. Choose the ground height maximizing vertex counts within ±2.5 cm across the sequence, then shift the whole motion to zero ground. This prevents frame-wise shifts from introducing jitter and avoids defining the plane solely by one extreme low point.

Complete original PDF figure A1.F5.
Figure 5: SMPL-X Foot Vertices for Ground-Contact Detection. This figure illustrates the selected foot vertices on the SMPL-X model used to detect ground contact. Blue and green points denote the left heel and left toe, while yellow and red represent the right heel and right toe, respectively. The remaining foot vertices are shown in light-gray. The clusters of colored points correspond to the specific parts of the foot that are used to check for contact with the ground, making the process more accurate and robust than using a single point.
Table 6: SMPL-X foot vertex indices used for ground–contact detection.
Region Vertex indices
Left heel 8888, 8889, 8891, 8909, 8910, 8911, 8913, 8914, 8915, 8916, 8917,
8918, 8919, 8920, 8921, 8922, 8923, 8924, 8925, 8929, 8930, 8934
Left toe 5773, 5781, 5782, 5791, 5793, 5805, 5808, 5816, 5817, 5830, 5831,
5859, 5860, 5906, 5907, 5908, 5909, 5912, 5914, 5915, 5916, 5917
Right heel 8676, 8677, 8679, 8697, 8698, 8699, 8701, 8702, 8703, 8704, 8705,
8706, 8707, 8708, 8709, 8710, 8711, 8712, 8713, 8714, 8715, 8716
Right toe 8467, 8475, 8476, 8485, 8487, 8499, 8502, 8510, 8511, 8524, 8525,
8553, 8554, 8600, 8601, 8602, 8603, 8606, 8608, 8609, 8610, 8611
Equation 11
g∗=arg⁡max𝑔∑f=1F∑v∈𝒱[|hvf−g|<δ],g^{*}=\underset{g}{\arg\max}\sum_{f=1}^{F}\sum_{v\in\mathcal{V}}\mathbf{1}\!\left[\,|h_{v}^{f}-g|<\delta\,\right], (11)
Complete original PDF figure A1.F6.
Figure 6: Procedure of Majority-Vote Ground Estimation. (1) shows how the foot-height distribution is obtained across all frames of a single human motion trajectory, and (2) shows how the ground height is selected from this distribution by majority vote.

A.1.3 Filtering Motion Data by Physical Information

Require root jerk below 50 m/s³, mean maximum foot-region contact ratio above 0.6, pelvis height between 0.6 and 1.5 m, and horizontal pelvis/spine1 distances from the support hull below 6/11 cm. Airborne phases can survive if surrounding contact keeps the clip score high enough.

The appendix defines the support hull from projected left/right foot and ankle joints. Thresholds follow manual inspection of Humanoid-X metric distributions; jerk and maximum height correspond to the 94th/99th percentiles. Other thresholds also exclude motions outside the flat-ground scope. These are selected heuristics, not universally validated physical feasibility boundaries.

Equation 12
Foot contact score=1T​∑t=1Tmax⁡(ctL​H,ctL​T,ctR​H,ctR​T),\text{Foot contact score}=\frac{1}{T}\sum_{t=1}^{T}\max\left(c^{LH}_{t},c^{LT}_{t},c^{RH}_{t},c^{RT}_{t}\right), (12)
Complete original PDF figure A1.F7.
Figure 7: Examples of Calculating the Ground-Contact Ratio. This figure illustrates how the ground-contact ratio is computed, which is used in the foot-contact score for physics-aware curation and in the ground and skating losses for PhySINK.
Table 7: Physics-aware data filtering metrics and thresholds.
Metric Threshold
Root jerk <50<50 m/s3
Foot contact score >0.6>0.6
Minimum pelvis height >0.6>0.6 m
Maximum pelvis height <1.5<1.5 m
Pelvis distance to base of support <6<6 cm
Spine1 distance to base of support <11<11 cm
Complete original PDF figure A1.F8.
Figure 8: Distribution of Each Curation Metric over Humanoid-X. The dashed line marks each threshold with its percentile, and motions in the shaded region marked with a cross are discarded.

A.2 Dataset Composition and Statistics

Table 8 totals 76.01K clips, 7.88M frames and 72.96 hours across mocap and video sources. Motion-X video contributes 31.98 hours and AMASS 20.86 hours; smaller sources broaden coverage. Figure 9 retains both complete source-distribution charts.

The LocoMuJoCo row lists 0.93M frames but 0.86 hours, inconsistent with 30 Hz and the total; that printed row is retained without guessing a correction. The aggregate duration also implies variable retained clip lengths, so 76,000 should not be multiplied by four seconds as if every surviving clip had identical duration.

Complete original PDF figure A1.F9. (a) (b)
Figure 9: Comparison of Motion Distributions. (a) Comparison between MoCap and Video sources in PHUMA. (b) Comparison with other existing datasets.
Table 8: Composition of the PHUMA dataset. A summary of the number of clips and duration for each sub-dataset, categorized by source and scene. PHUMA aggregates these diverse sub-datasets, resulting in 73.0 hours of physically reliable motion clips.
Dataset # Clip # Frame Duration Source Scene
LocoMuJoCo [21] 0.78K 0.93M 0.86h Motion Capture Indoor
GRAB [85] 1.73K 0.20M 1.88h Motion Capture Indoor
EgoBody [20] 2.12K 0.24M 2.19h Motion Capture Indoor
LAFAN1 [12] 2.18K 0.26M 2.40h Motion Capture Indoor
AMASS [13] 21.73K 2.25M 20.86h Motion Capture Indoor
HAA500 [29] 1.76K 0.11M 1.01h Human Video Outdoor
Motion-X Video [27] 33.04K 3.45M 31.98h Human Video Outdoor
HuMMan [30] 0.50K 0.05M 0.47h Human Video Indoor
AIST [31] 1.75K 0.18M 1.66h Human Video Indoor
IDEA400 [27] 9.94K 0.98M 9.10h Human Video Indoor
PHUMA Video 0.50K 0.06M 0.56h Human Video Outdoor
PHUMA 76.01K 7.88M 72.96h

Appendix B Motion Retargeting

The additional retargeting experiments compare morphology handling, offline throughput and initialization for a separate dynamics-aware optimizer.

B.1 Comparison of Retargeting Methods

GMR uses a height-derived factor and hand-tuned per-limb scaling. PhySINK instead optimizes SMPL shape to a shared robot T-pose and applies the original joint angles to that shape. Figures 10–11 show selected walking and extreme-height cases; they illustrate the mechanism without proving morphology-independent performance for every subject.

On curated sources excluding pre-retargeted datasets, PhySINK versus GMR gives 89.9% versus 84.0% PHUMA Test success and 81.7% versus 75.2% Unseen Video success. Unseen stationary performance is lower (97.4% versus 99.1%), and vertical performance ties, so the aggregate gain is not a win in every cell.

Complete original PDF figure A2.F10.
Figure 10: Qualitative Comparison of Retargeting Methods. This figure provides a visual comparison of human motion retargeted to a humanoid robot using the IK, SINK, and PhySINK methods. The top row shows the original human motion from the SMPL model, while the rows below show the resulting motions for each retargeting method.
Complete original PDF figure A2.F11.
Figure 11: Qualitative Comparison of Motion Retargeting: GMR vs. PhySINK. Comparison showing the limitations of GMR when handling extreme human heights. The top row illustrates retargeting from a short subject, where GMR causes excessive motion compression and crouching. The bottom row illustrates retargeting from a tall subject, where GMR generates infeasible targets, resulting in artifacts like joint locking (e.g., rigid knees) and loss of ground contact. PhySINK maintains kinematic feasibility in both cases.
Table 9: Imitation Performance: GMR vs. PhySINK Retargeting on Unitree G1. We evaluate the imitation performance of the MaskedMimic policy trained using datasets retargeted with GMR and PhySINK on the Unitree G1.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
GMR 84.0 92.1 77.8 77.1 89.1 75.2 99.1 77.8 61.7 52.7
PhySINK 89.9 94.2 87.6 84.2 91.8 81.7 97.4 86.7 61.7 71.4

B.2 Computational Cost

On one RTX 3090, throughput is 38.4 fps for IK, 25.3 for GMR, 0.95 for SINK and 0.92 for PhySINK. The latter optimize whole trajectories to handle temporal noise, making them suitable here for offline dataset preparation rather than real-time retargeting.

Table 10: Retargeting Throughput. Frames per second of each retargeting method on a single RTX 3090 GPU.
Retarget FPS
IK 38.4
GMR 25.3
SINK 0.95
PhySINK 0.92

B.3 PhySINK as Initialization for Dynamics-Aware Retargeting

Replace GMR initialization in SPIDER with PhySINK while retaining sampling-based optimization. Open-loop Unseen Video success rises from 4.6% to 10.7%. This supports a better starting point, while the low absolute success underscores that kinematic improvements do not solve dynamics by themselves.

Table 11: PhySINK as Initialization for Sampling-Based Retargeting. Open-loop success rate (%) on the Unseen Video motions when the sampling-based optimization of SPIDER starts from GMR, as in SPIDER, or from PhySINK.
Retarget Open-loop SR (%)
GMR ++ Sampling (SPIDER) 4.6
PhySINK ++ Sampling 10.7

Appendix C Motion Imitation Learning

The policy details specify what each controller observes, optimizes and randomizes. Preserved tables provide complete dimensions, reward formulas, hyperparameters and randomization ranges.

C.1 Observation Space Compositions

Both controllers combine current proprioception with reference-motion information, but their representations and intended deployment differ.

C.1.1 MaskedMimic

MaskedMimic observes root height and body position, rotation, velocity and angular velocity. Its goal contains absolute/root-relative and current-state-relative targets for 15 future timesteps. G1/H1-2 have 33/25 modeled bodies and total observation dimensions 9,898/7,498.

Table 12: Observation Space Dimensions of MaskedMimic.
Dimension
State G1 H1-2
(a) Proprioceptive State
Root height 1 1
Body position 32×332\times 3 24×324\times 3
Body rotation 33×633\times 6 25×625\times 6
Body velocity 33×333\times 3 25×325\times 3
Body angular velocity 33×333\times 3 25×325\times 3
(b) Goal State
Relative body position 33×15×333\times 15\times 3 25×15×325\times 15\times 3
Absolute body position 33×15×333\times 15\times 3 25×15×325\times 15\times 3
Relative body rotation 33×15×633\times 15\times 6 25×15×625\times 15\times 6
Absolute body rotation 33×15×633\times 15\times 6 25×15×625\times 15\times 6
Time 33×15×133\times 15\times 1 25×15×125\times 15\times 1
Total dim 9898 7498

C.1.2 BeyondMimic

BeyondMimic observes previous action, base angular velocity and joint position/velocity, plus instantaneous reference joints and torso orientation. Twenty reference frames spaced across 95 control steps (1.9 seconds at 50 Hz) form a 20×78 tensor, encoded to 128 dimensions; the full input is 282 dimensions.

Future targets raise 0.5 m success from 47.15% to 78.77% in IsaacSim and 46.91% to 76.02% in MuJoCo on PHUMA Test. Tracking errors in Table 14 are averaged only over successful motions. Acceleration error slightly increases, so the larger success set does not imply improvement in every conditional error metric.

Table 13: Observation Space Dimensions of BeyondMimic.
State Dimension (G1)
(a) Proprioceptive State
Previous action 29
Base angular velocity 3
DoF position 29
DoF velocity 29
(b) Goal State
(b.1) Instantaneous
DoF position 29
DoF velocity 29
Torso rotation 6
(b.2) Future Motion Targets (motion-encoder input)
Root height 20×120\times 1
Root roll/pitch 20×220\times 2
Base linear velocity 20×320\times 3
Base yaw rate 20×120\times 1
DoF position 20×2920\times 29
Local key-body position 20×14×320\times 14\times 3
Total dim 282
Table 14: Sim-to-Sim Results of PHUMA-trained BeyondMimic Policies. This table demonstrates the effect of the future-trajectory observation on the policy for the Unitree G1, evaluated on the PHUMA Test and Unseen Video splits. SR is the whole-body success rate at 0.5​m0.5\,\text{m} (%); Eg-mpjpeE_{\text{g-mpjpe}}, El-mpjpeE_{\text{l-mpjpe}}, EaccE_{\text{acc}}, and EvelE_{\text{vel}} are tracking errors (mm), following the convention of [37, 9, 11] and averaged over successful motions.
Future IsaacSim MuJoCo
SR ↑\uparrow Eg-mpjpe↓E_{\text{g-mpjpe}}\downarrow El-mpjpe↓E_{\text{l-mpjpe}}\downarrow Eacc↓E_{\text{acc}}\downarrow Evel↓E_{\text{vel}}\downarrow SR ↑\uparrow Eg-mpjpe↓E_{\text{g-mpjpe}}\downarrow El-mpjpe↓E_{\text{l-mpjpe}}\downarrow Eacc↓E_{\text{acc}}\downarrow Evel↓E_{\text{vel}}\downarrow
(a) PHUMA test
✗ 47.15 115.78 32.70 1.15 4.29 46.91 156.64 32.92 1.22 4.36
✓ 78.77 112.00 28.28 1.18 4.14 76.02 142.79 28.98 1.24 4.18
(b) Unseen Video
✗ 47.82 123.71 31.93 1.07 4.93 47.42 122.89 30.60 1.09 4.85
✓ 71.63 121.69 24.87 1.11 4.78 70.44 122.68 23.79 1.14 4.73

C.2 Reward Function

Tracking rewards encourage agreement with the target, while regularization discourages unstable or abrupt actions. The exact formulas and weights are preserved in Tables 15–16.

C.2.1 MaskedMimic

MaskedMimic rewards global body positions, root height, rotations and linear/angular velocities with exponential error functions. Power consumption is penalized, and the added action-change squared penalty has weight −0.2 to encourage smooth control.

Table 15: Reward Function Terms of MaskedMimic.
Term Expression Weight
(a) Task
Global body position exp(−100⋅∥pt−p^t∥22)\exp(-100\cdot\lVert p_{t}-\hat{p}_{t}\rVert^{2}_{2}) 0.5
Root height exp(−100⋅(htroot−h^troot)2)\exp(-100\cdot(h^{\text{root}}_{t}-\hat{h}^{\text{root}}_{t})^{2}) 0.2
Global body rotation exp(−10⋅∥θt⊖θ^t∥22)\exp(-10\cdot\lVert\theta_{t}\ominus\hat{\theta}_{t}\rVert^{2}_{2}) 0.3
Global body velocity exp(−0.5⋅∥vt−v^t∥22)\exp(-0.5\cdot\lVert v_{t}-\hat{v}_{t}\rVert^{2}_{2}) 0.1
Global body angular velocity exp(−0.1⋅∥ωt−ω^t∥22)\exp(-0.1\cdot\lVert\omega_{t}-\hat{\omega}_{t}\rVert^{2}_{2}) 0.1
(b) Regularization
Power consumption ∥F⊙q˙∥1\lVert F\odot\dot{q}\rVert_{1} -1e-05
Action rate ∥at−at−1∥22\lVert a_{t}-a_{t-1}\rVert^{2}_{2} -0.2

C.2.2 BeyondMimic

BeyondMimic combines global reference tracking with body targets re-anchored to the robot’s root after factoring out reference horizontal translation and heading. Regularizers penalize action changes, joint-position limit violations and undesired contacts. These controller rewards are distinct from the offline PhySINK losses.

Table 16: Reward Function Terms of BeyondMimic.
Term Expression Weight
(a) Motion Tracking
Reference body position exp(−∥ptref−p^tref∥22/ 0.32)\exp(-\lVert p^{\text{ref}}_{t}-\hat{p}^{\text{ref}}_{t}\rVert^{2}_{2}\,/\,0.3^{2}) 0.5
Reference body rotation exp(−∥θtref⊖θ^tref∥22/ 0.42)\exp(-\lVert\theta^{\text{ref}}_{t}\ominus\hat{\theta}^{\text{ref}}_{t}\rVert^{2}_{2}\,/\,0.4^{2}) 0.5
Global body linear velocity exp(−1B∑b∥vtb−v^tb∥22/ 1.02)\exp\!\left(-\tfrac{1}{B}\sum_{b}\lVert v^{b}_{t}-\hat{v}^{b}_{t}\rVert^{2}_{2}\,/\,1.0^{2}\right) 1.0
Global body angular velocity exp(−1B∑b∥ωtb−ω^tb∥22/π2)\exp\!\left(-\tfrac{1}{B}\sum_{b}\lVert\omega^{b}_{t}-\hat{\omega}^{b}_{t}\rVert^{2}_{2}\,/\,\pi^{2}\right) 1.0
Relative body position exp(−1B∑b∥ptb−p^tb∥22/ 0.32)\exp\!\left(-\tfrac{1}{B}\sum_{b}\lVert p^{b}_{t}-\hat{p}^{b}_{t}\rVert^{2}_{2}\,/\,0.3^{2}\right) 1.0
Relative body rotation exp(−1B∑b∥θtb⊖θ^tb∥22/ 0.42)\exp\!\left(-\tfrac{1}{B}\sum_{b}\lVert\theta^{b}_{t}\ominus\hat{\theta}^{b}_{t}\rVert^{2}_{2}\,/\,0.4^{2}\right) 1.0
(b) Regularization
Action rate ∥at−at−1∥22\lVert a_{t}-a_{t-1}\rVert^{2}_{2} -0.1
DoF position-limit violation ∑j(ReLU⁡(qtj−qmax,0.9j)+ReLU⁡(qmin,0.9j−qtj))\sum_{j}\big(\,\mathrm{ReLU}(q^{j}_{t}-q^{j}_{\max,0.9})+\mathrm{ReLU}(q^{j}_{\min,0.9}-q^{j}_{t})\,\big) -100
Undesired contacts ∑b[∥Ftb∥2>1N]\sum_{b}\mathbf{1}\!\left[\,\lVert F^{b}_{t}\rVert_{2}>1\,\text{N}\,\right] -0.5

C.3 PPO Hyperparameters

Both configurations use 8,192 environments. MaskedMimic uses Adam, a four-layer 512-dimensional transformer actor and a four-layer 1,024-wide critic. BeyondMimic uses AdamW, 768/512/256 MLPs, 24 rollout steps, eight epochs and a 60-hidden/128-output motion encoder. The complete source values remain in Tables 17–18, including the unusual printed MaskedMimic “Init noise std” of −2.9; its parameterization is not clarified here.

Table 17: PPO Hyperparameter Values for Model Training of MaskedMimic.
Hyperparameter Value
Optimizer Adam
Num envs 8192
Mini Batches 32
Learning epochs 1
Entropy coefficient 0.0
Value loss coefficient 0.5
Clip param 0.2
Max grad norm 50.0
Init noise std -2.9
Actor learning rate 2e-5
Critic learning rate 1e-4
GAE decay factor(λ\lambda) 0.95
GAE discount factor(γ\gamma) 0.99
Actor Transformer dimension 512
Actor layers 4
Actor heads 4
Critic MLP size [1024, 1024, 1024, 1024]
Activation ReLU
Table 18: PPO Hyperparameter Values for Model Training of BeyondMimic.
Hyperparameter Value
Optimizer AdamW
Num envs 8192
Steps per env 24
Mini batches 4
Learning epochs 8
Entropy coefficient 0.005
Value loss coefficient 1.0
Clip param 0.2
Max grad norm 1.0
Init noise std 1.0
Actor learning rate 1e-5
Critic learning rate 1e-5
GAE decay factor(λ\lambda) 0.95
GAE discount factor(γ\gamma) 0.99
Actor MLP size [768, 512, 256]
Critic MLP size [768, 512, 256]
Motion encoder hidden dim 60
Motion encoder output dim 128
Activation SiLU

C.4 Domain Randomization

Domain randomization targets the simulation-to-hardware gap during training rather than changing the reference-motion dataset.

C.4.1 BeyondMimic

BeyondMimic varies actuator gains, joint bias, base center of mass, friction/restitution, pushes, initial state and actor-observation noise. The critic receives noiseless observations. Table 19 preserves ranges and schedules; the paper states that randomization and noise are disabled during evaluation, so these evaluations are not a stress test across all sampled disturbances.

Table 19: Domain Randomization Ranges of BeyondMimic.
Parameter Range
P-gain scale sPs_{P} 𝒰⁡(0.9,1.1)\mathcal{U}(0.9,1.1)
D-gain scale sDs_{D} 𝒰⁡(0.9,1.1)\mathcal{U}(0.9,1.1)
DoF position bias 𝒰⁡(−0.025,0.025)​rad\mathcal{U}(-0.025,0.025)\,\text{rad}
Base CoM offset xx 𝒰⁡(−0.025,0.025)​m\mathcal{U}(-0.025,0.025)\,\text{m}
Base CoM offset y,zy,z 𝒰⁡(−0.05,0.05)​m\mathcal{U}(-0.05,0.05)\,\text{m}
Static friction 𝒰⁡(0.3,1.6)\mathcal{U}(0.3,1.6)
Dynamic friction 𝒰⁡(0.3,1.2)\mathcal{U}(0.3,1.2)
Restitution 𝒰⁡(0.0,0.5)\mathcal{U}(0.0,0.5)
Push interval 𝒰⁡(1.0,3.0)​s\mathcal{U}(1.0,3.0)\,\text{s}
Push linear velocity 𝒰⁡(±0.5,±0.5,±0.2)​m/s\mathcal{U}(\pm 0.5,\pm 0.5,\pm 0.2)\,\text{m/s}
Push angular velocity 𝒰⁡(±0.52,±0.52,±0.78)​rad/s\mathcal{U}(\pm 0.52,\pm 0.52,\pm 0.78)\,\text{rad/s}
Initial DoF position 𝒰⁡(−0.1,0.1)​rad\mathcal{U}(-0.1,0.1)\,\text{rad}
Initial root position 𝒰⁡(±0.05,±0.05,±0.01)​m\mathcal{U}(\pm 0.05,\pm 0.05,\pm 0.01)\,\text{m}
Initial root rotation 𝒰⁡(±0.1,±0.1,±0.2)​rad\mathcal{U}(\pm 0.1,\pm 0.1,\pm 0.2)\,\text{rad}
Initial root linear vel. 𝒰⁡(±0.1,±0.1,±0.05)​m/s\mathcal{U}(\pm 0.1,\pm 0.1,\pm 0.05)\,\text{m/s}
Initial root angular vel. 𝒰⁡(±0.1,±0.1,±0.1)​rad/s\mathcal{U}(\pm 0.1,\pm 0.1,\pm 0.1)\,\text{rad/s}
Obs. noise: base angular vel. (σ\sigma) 0.2​rad/s0.2\,\text{rad/s}
Obs. noise: DoF position (σ\sigma) 0.01​rad0.01\,\text{rad}
Obs. noise: DoF velocity (σ\sigma) 0.5​rad/s0.5\,\text{rad/s}
Obs. noise: motion ref orientation (σ\sigma) 0.050.05

Appendix D Experiment Details

Additional tests distinguish strict tracking, alternative reference construction, simulator transfer and falling frequency.

D.1 Self-Collected Video Dataset

The 504 self-recorded videos span 11 motion types. TRAM extracts human motion and PhySINK maps it to robot references. These are reported as unseen recordings, although reconstruction and retargeting introduce their own errors. Appendix D.4 addresses dependence on the chosen retargeter.

Complete original PDF figure A4.F12.
Figure 12: Overview of the Self-collected Data Pipeline. This figure illustrates the three main steps of our data collection pipeline: (left) a self-recorded video of a human motion, (center) the motion extracted using a video-to-motion model, and (right) the final motion retargeted to a humanoid robot.

D.2 Success Rate Threshold Analysis

Success depends strongly on tolerance. On G1 PHUMA Test, PHUMA/AMASS achieve 92.7%/76.2% at 0.15 m but 97.1%/90.2% at 0.5 m. On Unseen Video, the corresponding totals are 82.9%/70.2% and 93.7%/92.3%. The strict threshold separates tracking precision more clearly; changing the threshold also changes category rankings.

Table 20: Performance Comparison based on Success Threshold in PHUMA  Test.
Success Threshold=0.15m Success Threshold=0.5m
Dataset Hours Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
(a) G1
LaFAN1 2.4 46.1 66.1 36.2 24.0 42.5 74.8 87.8 69.2 47.1 72.6
AMASS 20.9 76.2 88.5 72.1 56.8 66.8 90.2 95.0 87.9 81.1 83.7
Humanoid-X 231.4 50.6 78.4 43.0 26.0 31.8 78.4 91.3 72.9 59.5 65.9
PHUMA 73.0 92.7 95.6 91.7 86.0 85.6 97.1 98.7 96.5 94.4 92.5
(b) H1-2
LaFAN1 2.4 62.0 79.3 54.7 26.6 58.9 70.8 92.4 66.7 56.4 68.2
AMASS 20.9 54.4 74.9 45.9 17.2 49.6 70.4 86.3 62.6 41.4 65.9
Humanoid-X 231.4 49.7 74.6 40.4 17.0 37.3 54.8 78.5 45.2 22.1 43.2
PHUMA 73.0 82.7 91.5 79.5 68.1 68.4 92.0 96.6 89.7 85.6 79.4
Table 21: Performance Comparison based on Success Threshold in Unseen Video.
Success Threshold=0.15m Success Threshold=0.5m
Dataset Hours Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
(a) G1
LaFAN1 2.4 28.4 46.9 28.4 19.6 10.5 78.2 85.5 70.8 76.3 80.8
AMASS 20.9 70.2 90.7 75.0 62.7 44.1 92.3 99.2 92.1 82.1 88.0
Humanoid-X 231.4 39.1 78.0 39.6 23.0 6.5 84.1 98.3 79.9 76.0 76.2
PHUMA 73.0 82.9 96.7 88.0 71.8 67.1 93.7 100.0 96.8 85.9 84.7
(b) H1-2
LaFAN1 2.4 70.8 92.4 66.7 56.4 68.2 85.5 97.5 79.0 77.5 90.0
AMASS 20.9 64.3 87.3 59.7 46.0 63.9 80.4 93.3 69.9 72.8 89.0
Humanoid-X 231.4 60.5 88.3 60.0 48.7 39.7 68.7 93.3 65.1 60.2 50.5
PHUMA 73.0 78.6 97.5 76.8 74.5 63.8 89.9 99.2 89.4 84.6 83.9

D.3 Cross-Simulator Generalization Analysis

Teacher/student policies using the KungfuBot2 reference are trained in Isaac Gym and transferred to MuJoCo. PHUMA improves aggregate teacher, student and transfer success. In MuJoCo, totals are 75.0%/70.0% versus AMASS 62.1%/64.3% on the two test sets. Some in-training category results tie or favor AMASS, and simulator transfer remains different evidence from hardware deployment.

Table 22: Imitation Performance of the KungfuBot Teacher Policy in Isaac Gym. We evaluate the imitation performance of KungfuBot teacher policy with unseen motions on the Unitree G1.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
AMASS 77.8 90.8 71.2 58.3 84.7 83.5 100.0 83.3 74.5 72.5
PHUMA 91.0 96.4 88.3 87.6 90.5 87.1 99.1 89.2 77.7 76.9
Table 23: Imitation Performance of the KungfuBot Student Policy in Isaac Gym. We evaluate the imitation performance of KungfuBot student policy with unseen motions on the Unitree G1.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
AMASS 66.6 86.7 58.8 41.8 68.4 67.7 97.4 68.0 59.6 37.4
PHUMA 82.9 93.5 78.5 74.2 81.7 73.8 97.4 76.9 63.8 47.3
Table 24: Sim-to-Sim transfer performance across motion dataset. We evaluate the zero-shot motion tracking success rate of policies trained on PHUMA and AMASS when transferred from the source simulator (Isaac Gym) to the target simulator (MuJoCo). The results are demonstrated on the G1 humanoid to assess robustness against domain shifts in physics engines.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
AMASS 62.1 81.4 54.2 38.8 64.3 64.3 86.2 68.5 54.3 37.4
PHUMA 75.0 87.6 69.3 61.6 76.3 70.0 87.9 78.8 59.6 38.5

D.4 Evaluation on GMR-Retargeted References

Retarget evaluation motions with GMR and use 0.5 m success, reporting mean ± standard deviation over three seeds. PhySINK-trained AMASS reaches 83.4±0.2%/83.7±0.4%; complete PHUMA reaches 95.5±0.2%/89.4±0.4%. Thus aggregate benefits persist with different references and after comparing against AMASS using the same improved retargeter. These uncertainties apply to this protocol, not automatically to every main-table result.

Table 25: Imitation Performance on GMR-Retargeted References: Retargeting Methods. Success rate (%) at the 0.5m threshold of policies trained on AMASS retargeted by each method, evaluated on the PHUMA Test and Unseen Video motions retargeted with GMR. Mean (±\pm std) over three seeds.
Retarget PHUMA Test (GMR) Unseen Video (GMR)
IK 80.3 (±\pm0.2) 81.6 (±\pm0.6)
GMR 80.0 (±\pm0.1) 81.1 (±\pm0.3)
SINK 81.6 (±\pm0.2) 79.8 (±\pm0.6)
PhySINK 83.4 (±\pm0.2) 83.7 (±\pm0.4)
Table 26: Imitation Performance on GMR-Retargeted References: Datasets. Success rate (%) at the 0.5m threshold of policies trained on each dataset, evaluated on the PHUMA Test and Unseen Video motions retargeted with GMR. Mean (±\pm std) over three seeds.
Dataset PHUMA Test (GMR) Unseen Video (GMR)
LaFAN1 76.5 (±\pm0.1) 81.6 (±\pm0.9)
Humanoid-X 77.7 (±\pm0.1) 79.6 (±\pm0.5)
AMASS (SINK) 81.6 (±\pm0.2) 79.8 (±\pm0.6)
AMASS (PhySINK) 83.4 (±\pm0.2) 83.7 (±\pm0.4)
PHUMA 95.5 (±\pm0.2) 89.4 (±\pm0.4)

D.5 Fall Rate of Humanoid-X and AMASS Policies

A fall means pelvis height below 0.3 m. Humanoid-X versus AMASS falls in 18.6% versus 5.7% of PHUMA Test episodes and 13.1% versus 4.2% of Unseen Video episodes. This motivates omitting Humanoid-X hardware deployment, while limiting the range of baselines tested on the real robot.

Table 27: Fall Rate in Simulation. Percentage of episodes in which the pelvis height drops below 0.3m, for policies trained on Humanoid-X and on AMASS retargeted with SINK. Lower is better.
Dataset PHUMA Test FR (%) ↓\downarrow Unseen Video FR (%) ↓\downarrow
Humanoid-X 18.6 13.1
AMASS (SINK) 5.7 4.2

Appendix E Ablation Studies

The ablations examine source modality, filtering, reconstruction quality and retargeting penalties. Several use different data subsets or evaluation references, so their totals should not be merged into one sequential improvement curve.

E.1 Mocap only and Video only data performance

Video-only training improves all listed tracking categories over mocap-only, while both yield high kinematic quality after retargeting. The video subset is larger and more varied, so this does not isolate modality from scale and coverage.

Adding all video to a mocap base gives the best 0.5 m GMR-reference results, 94.6±0.1%/90.2±0.9%. The trend is not monotonic: adding 33% video lowers Unseen Video success from 86.0% to 80.1% before larger additions improve it. Table 30 retains this exception.

Table 28: Performance Evaluation of PHUMA based on Data Source (MoCap vs. Video). We present a quantitative comparison evaluating the performance achieved using PHUMA data derived from motion capture (MoCap) versus video, concluding that both sources offer competitive results.
Motion Fidelity (%) Joint Feasibility (%) Non-Floating (%) Non-Penetration (%) Non-Skating (%)
MoCap 96.7 100.0 99.9 94.3 92.1
Video 93.8 100.0 99.9 98.0 88.3
Table 29: Imitation Performance of PHUMA based on Data Source (MoCap vs. Video). We evaluate the imitation performance of MaskedMimic policy trained with PHUMA data derived from motion capture (MoCap) versus video on the Unitree G1.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
MoCap 75.2 88.2 68.4 49.0 86.9 73.0 96.6 73.9 56.4 58.2
Video 85.7 92.9 81.8 75.5 89.5 76.2 100.0 76.4 59.6 62.6
Table 30: Imitation Performance When Adding Video to Mocap. Success rate (%) at the 0.5m threshold on GMR-retargeted references, for policies trained on the mocap-sourced motions plus a growing fraction of the video-sourced motions. Mean (±\pm std) over three seeds.
Data Source PHUMA Test (GMR) Unseen Video (GMR)
MoCap Only 82.2 (±\pm0.2) 86.0 (±\pm0.2)
++ Video (33%) 82.1 (±\pm0.1) 80.1 (±\pm0.8)
++ Video (66%) 92.8 (±\pm0.1) 86.7 (±\pm0.5)
++ Video (100%) 94.6 (±\pm0.1) 90.2 (±\pm0.9)

E.2 Physics-based Filtering

Filtering studies separate retained volume, retargeted kinematic quality and downstream controller performance. They answer related but different questions.

E.2.1 Data Distribution Based on Physics-based Filtering

Sequential jerk, contact, height and support filters leave 27.1% in the illustrated flow. The prose describes Humanoid-X, whereas Figure 13’s caption and diagram name Humanoid-X plus LaFAN1/LocoMuJoCo. This source-scope discrepancy is preserved, so that percentage should not be treated as an exact audited yield for every source.

Complete original PDF figure A5.F13.
Figure 13: Dataset Statistics After Physics-based Filtering. Distribution of motion sequences after applying physics-based filtering to the combined Humanoid-X, LaFAN1, and LocoMuJoCo datasets.

E.2.2 PhySINK’s Robustness to Noisy Motion Sources

PhySINK maintains 100% joint feasibility across listed input variants, but fidelity improves from 70.6% on raw data to 94.8% with all filters. Good retargeting therefore does not eliminate the need for clean inputs. The prose describes individual filters, consistent with nonmonotonic retained hours, while the caption says “sequentially”; raw hours also appear as 237.2 here versus 231.4 elsewhere. These reported values are retained without conflating the protocols.

Table 31: PhySINK Retargeting Robustness to Noisy Motion Sources. This figure presents an experiment to evaluate how robustly the PhySINK method retargets various noisy Humanoid-X motion sources. The six distinct motion source groups used for retargeting are compared: (1) Original Humanoid-X, and Humanoid-X motion sources sequentially refined by applying (2) Jerk filtering, (3) Foot Contact filtering, (4) Height filtering, (5) BoS filtering, and (6) All filtering (PHUMA).
Motion Source Hours Motion Fidelity (%) Joint Feasibility (%) Non-Floating (%) Non-Penetration (%) Non-Skating (%)
Humanoid-X 237.2 70.6 100.0 98.7 92.2 90.1
Jerk Filter 141.1 76.9 100.0 99.7 96.5 92.0
Foot Contact Filter 123.1 77.1 100.0 99.8 96.8 90.6
Height Filter 135.7 87.0 100.0 99.4 94.7 89.8
BoS Filter 110.3 90.7 100.0 99.4 94.9 90.0
All Filter 62.2 94.8 100.0 99.9 96.7 89.7

E.2.3 Leave-One-Out Ablation of Curation Filters

With PhySINK and 0.5 m GMR references, all filters yield 94.6±0.1%/90.2±0.9%. Removing any filter increases volume but reduces success; removing foot contact gives the largest decline, to 87.5±0.1%/80.0±0.5%. This supports contact-based selection within the chosen flat-ground criteria.

Table 32: Leave-One-Out Ablation of Curation Filters. Success rate (%) at the 0.5m threshold on GMR-retargeted references, for policies trained on motions curated with all filters or with one filter removed, all retargeted with PhySINK. Hours give the curated volume. Mean (±\pm std) over three seeds.
Curation Filter Hours PHUMA Test (GMR) Unseen Video (GMR)
All Filters 62 94.6 (±\pm0.1) 90.2 (±\pm0.9)
No Jerk Filter 75 91.8 (±\pm0.1) 89.5 (±\pm0.3)
No FC Filter 77 87.5 (±\pm0.1) 80.0 (±\pm0.5)
No Height Filter 73 89.2 (±\pm0.3) 85.6 (±\pm0.1)
No BoS Filter 88 89.5 (±\pm0.1) 87.4 (±\pm0.7)

E.2.4 Sensitivity to the Video Reconstruction Model

GVHMR raises curation acceptance from TRAM’s 41.9% to 77.4%, yet PhySINK still improves joint feasibility and contact metrics over SINK. Non-skating is 93.8% for TRAM/PhySINK and 86.6% for GVHMR/PhySINK, so better acceptance is not improvement on every downstream quality metric. The comparisons concern the accepted subsets, which need not contain identical motions.

Table 33: Sensitivity to the Video Reconstruction Model. Curation acceptance rate on the Unseen Video motions and retargeting quality of SINK and PhySINK, for motions reconstructed with TRAM and GVHMR.
Reconstruction Acceptance Retarget Retargeting Quality
Rate (%) MF (%) JF (%) NF (%) NP (%) NS (%)
TRAM 41.9 SINK 95.9 63.6 100.0 27.2 40.1
PhySINK 95.9 100.0 100.0 99.7 93.8
GVHMR 77.4 SINK 95.6 78.7 99.9 27.3 64.8
PhySINK 95.6 100.0 100.0 100.0 86.6

E.3 Impact of Motion Retargeting Quality on Policy Performance

Shape-adaptive SINK variants outperform IK/GMR in these aggregate policy tests. Adding physical penalties gives smaller, nonmonotonic gains: joint-feasibility-only lowers totals; grounding gives 90.0% PHUMA Test, while full PhySINK gives 89.9% there and the best 81.7% Unseen Video. The authors interpret fidelity as especially important, but these coupled-method comparisons do not establish that every other artifact is causally unimportant.

Table 34: Ablation Studies of Imitation Performance on Retargeting Loss. We evaluate the imitation performance of MaskedMimic policies trained with and without the physical constraint loss (Table 1) using the Unitree G1 robot.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
IK 70.5 85.9 63.3 47.6 77.3 68.5 96.6 72.4 43.6 49.5
GMR 84.0 92.1 77.8 77.1 89.1 75.2 99.1 77.8 61.7 52.7
SINK 89.1 94.0 86.0 84.9 90.7 79.0 100.0 81.8 62.8 62.6
+ Joint Feasibility Loss 87.0 92.1 83.6 79.8 90.9 78.6 94.8 84.7 56.4 67.0
+ Grounding Loss 90.0 93.5 87.9 85.6 92.0 80.4 98.3 86.2 60.6 64.8
+ Skating Loss 89.9 94.2 87.6 84.2 91.8 81.7 97.4 86.7 61.7 71.4

E.4 Leave-One-Out Ablation of PhySINK Losses

The separate leave-one-loss-out test uses 0.5 m GMR references and three seeds. Full PhySINK is best; dropping grounding lowers success to 90.5±0.1%/84.5±0.8% and non-penetration to 14.6%. Dropping skating reduces non-skating from 89.7% to 53.6%. This demonstrates meaningful contact penalties under that protocol, despite diminishing or mixed gains in the earlier cumulative ablation.

Table 35: Leave-One-Out Ablation of PhySINK Losses. Retargeting quality, and success rate (%) at the 0.5m threshold on GMR-retargeted references, with each loss removed from the full objective. Mean (±\pm std) over three seeds.
Retarget Retargeting Quality Motion Tracking Performance
MF (%) JF (%) NF (%) NP (%) NS (%) PHUMA Test (GMR) Unseen Video (GMR)
PhySINK 94.8 100.0 99.9 96.8 89.7 94.6 (±\pm0.1) 90.2 (±\pm0.9)
No Joint Feasibility Loss 91.7 95.8 99.9 96.4 89.4 91.0 (±\pm0.1) 85.8 (±\pm0.6)
No Grounding Loss 91.8 100.0 96.4 14.6 91.7 90.5 (±\pm0.1) 84.5 (±\pm0.8)
No Skating Loss 94.9 100.0 99.9 97.2 53.6 94.2 (±\pm0.1) 88.2 (±\pm0.8)

Appendix F Pelvis-Only Path Following Control Performance

Distill MaskedMimic students whose motion goal is only pelvis position/rotation. PHUMA improves aggregate path-following success on both robots: G1 reaches 84.5%/74.6% versus AMASS 60.5%/54.8%, and H1-2 reaches 73.9%/78.1% versus 60.4%/72.3%. Figure 14 preserves the running example and target path. This reduced goal does not mean the controller lacks proprioception, nor does the example establish arbitrary-path generalization.

Table 36: Pelvis path following peformance across motion dataset. We evaluate the success rate of pelvis path-following control for policies trained on the AMASS and PHUMA datasets across various pelvis trajectories from the PHUMA Test and Unseen Video.
PHUMA Test Unseen Video
Dataset Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
(a) G1
AMASS 60.5 85.6 60.1 51.4 66.5 54.8 83.6 66.5 33.0 27.5
PHUMA 84.5 94.6 86.1 83.7 90.2 74.6 98.3 83.3 54.3 57.1
(a) H1-2
AMASS 60.4 84.0 62.8 43.6 78.7 72.3 96.6 77.3 52.1 72.5
PHUMA 73.9 91.2 76.5 66.9 84.8 78.1 96.6 77.8 60.6 78.0
Complete original PDF figure A6.F14.
Figure 14: Path following on running motion. We visualize the robot’s trajectory in a running motion. The target pelvis path is visualized with a green line. Top row presents results from a policy trained on AMASS, while bottom row presents results from a policy trained on PHUMA.

Appendix G Results on H1-2 Robot

H1-2 results test a second morphology in simulation. The real-world results in the main text remain G1-only.

G.1 PhySINK Ablation on H1-2

SINK alone has only 15.3% joint feasibility on H1-2. PhySINK raises this to 99.9%, with 93.9% fidelity, 97.7% non-floating, 99.7% non-penetration and 87.7% non-skating. The large feasibility change highlights morphology dependence, while remaining artifacts preclude claiming perfect retargeting.

Table 37: Quantitative comparison and ablation study of retargeting methods on H1-2. Progressive impact of adding each physical constraint loss.
Motion Fidelity (%) Joint Feasibility (%) Non-Floating (%) Non-Penetration (%) Non-Skating (%)
IK 36.3 80.9 57.7 45.2 56.1
GMR 58.0 71.8 6.0 100.0 62.3
SINK 93.9 15.3 42.2 81.4 47.9
+ Joint Feasibility Loss 94.0 99.5 44.4 79.9 50.7
+ Grounding Loss 93.9 99.9 99.8 98.1 49.3
+ Skating Loss = PhySINK 93.9 99.9 97.7 99.7 87.7

G.2 Retargeting Method Comparison on H1-2

AMASS retargeted with PhySINK yields the best aggregate H1-2 success, 64.3%/72.4%. Individual horizontal categories favor GMR on PHUMA Test (65.9% versus 55.9%) and SINK on Unseen Video (63.9% versus 63.1%). The aggregate gain does not establish category-wise dominance.

Table 38: Motion tracking performance across retargeting approaches on H1-2. Success rates using two test sets.
PHUMA Test Unseen Video
Retarget Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
IK 45.3 70.9 35.7 15.2 35.0 54.2 78.0 60.7 30.1 28.6
GMR 56.6 75.7 48.1 24.8 65.9 64.3 80.2 65.5 52.1 53.8
SINK 54.4 74.9 45.9 17.2 49.6 64.3 87.3 59.7 46.0 63.9
PhySINK 64.3 83.6 57.0 27.7 55.9 72.4 99.2 66.3 57.4 63.1

G.3 Dataset Effectiveness on H1-2

PHUMA yields 82.7%/78.6% aggregate H1-2 success versus AMASS 54.4%/64.3%. However, Unseen Video horizontal success is 63.8%, below LaFAN1’s 68.2% and slightly below AMASS’s 63.9%. Table 39 therefore qualifies the main text’s claim that PHUMA wins across every category on both humanoids.

Table 39: Motion tracking performance across datasets on H1-2. Success rates evaluated on two test sets.
PHUMA Test Unseen Video
Dataset Hours Total Stationary Angular Vertical Horizontal Total Stationary Angular Vertical Horizontal
LaFAN1 2.4 62.0 79.3 54.7 26.6 58.9 70.8 92.4 66.7 56.4 68.2
AMASS 20.9 54.4 74.9 45.9 17.2 49.6 64.3 87.3 59.7 46.0 63.9
Humanoid-X 231.4 49.7 74.6 40.4 17.0 37.3 60.5 88.3 60.0 48.7 39.7
PHUMA 73.0 82.7 91.5 79.5 68.1 68.4 78.6 97.5 76.8 74.5 63.8
Concept demonstration

Open image ↗