MotionPersonaX

Locomotion generated by a controller trained on MotionPersona and MotionPersonaX. Each of the 44 characters is conditioned on the persona and body shape of one annotated participant and follows the same straight path; the formation spreads because characters with longer legs travel faster.

MotionPersonaX is a cross-body extension of the MotionPersona motion-capture dataset. Every annotated clip of MotionPersona is retargeted onto 128 SMPL-X bodies, the 48 captured participants and 80 synthetic bodies that tile height and girth, yielding 328,960 clips (about 4,200 h) in which the person who performed the motion and the body that carries it vary independently by construction. Retargeting is performed by an explicit optimisation that preserves the timing of the source and adapts posture, contacts and balance to the target body.

[homepage]  [arxiv]  [data preview]  [code]

Source clips	2,570 clips of the 44 annotated MotionPersona participants (9 contexts × 7 locomotion types)
Target bodies	128: 48 captured participants + 80 grid bodies (10 heights from 1.05 to 1.95 m × 8 girth levels)
Clips	328,960 = 2,570 × 128 (about 4,200 h; 453.5 M frames)
Frame rate	30 Hz (see Section 4.4 for exceptions)
Representations	SMPL-X parameters and BVH on the SMPL-X skeleton of each target body
Size	about 215 GB (one archive per body and representation)

Each participant is captured in exactly one body, their own, so the captured data occupy only the diagonal of the persona × body grid. MotionPersonaX fills all 44 × 128 cells.

1. Target bodies

The 128 target bodies are defined by 10-dimensional SMPL-X shape coefficients (smpl_skeleton/shapes/, bodies.csv).

• Captured bodies (48). The bodies of the MotionPersona participants, with the same identifiers (p01–p44, b15, b27, b33, b34). The four b bodies have no persona annotation and serve as targets only.
• Grid bodies (80). Synthetic bodies grid_h<i>_g<j> that tile ten heights (i = 0…9, 1.05 to 1.95 m in steps of 0.1 m) against eight girth levels (j = 0…7), so that the body range extends beyond the captured participants.

For each body, bodies.csv lists height, mass, leg length, hip width and hip height, computed from the SMPL-X mesh.

One clip on its source body (left) and on eight target bodies of increasing height, mass and hip width. All rows are data from this dataset.

2. Retargeting

Retargeting a clip onto another body requires a definition of the same motion on that body: which properties of the source are kept, and which change with the body. MotionPersonaX takes this definition from measurements on the captured data (Section 2.1) and realises it with an explicit optimisation (Section 2.2).

2.1 Measurement: what changes with the body

Because every participant performs the same grid of contexts and locomotion types, each of 27 gait descriptors (timing, space, limbs and trunk; lengths normalised by leg length, hip width or shoulder width) can be decomposed into a participant effect, a context effect and their interaction. The participant effect is large and stable: it exceeds the context effect by an order of magnitude for speed, jerk and sway, with split-half reliability of 0.57–0.96. Body measures, however, explain only a small part of it. Each participant effect is regressed on nine body measures with leave-one-participant-out prediction, and an association is attributed to the body only if it survives an age control, keeps its sign after partialling out age, has the same sign in all nine contexts, and is confirmed in direction by twelve metadata-matched adult pairs that share all three persona attributes but differ in body.

Two associations pass, both on trunk lean: wider hips go with more forward lean and more side lean.

The two associations assigned to body shape, measured on the captured data. Grey: each adult participant's effect on trunk lean (the participant term of a participant × context decomposition) against hip width; blue: hip-width quartile means ± SE; dashed: fit over all 39 adults with bootstrap 95 % band. Wider hips go with more forward lean (+4.5° per 2 cm over all adults, +2.6° within the 12 metadata-matched pairs that share all three persona attributes but differ in body) and more side lean (+1.2° and +1.5° per 2 cm).

The retargeter implements these two associations with slopes of k_f = 150° and k_s = 65° per metre of hip width, between the within-pair and the population estimates and inside the 95 % intervals of the pair slopes (41–249° and 19–135° per metre). No other descriptor passes all tests: no reliable body association is detected for cadence; the apparent ones for leg-normalised step length, knee range, foot clearance and joint speed vanish under the age control; and hand distance scales with shoulder width. These descriptors are therefore preserved from the source rather than adapted. Within the descriptors and the statistical resolution of this study, a variance-weighted shape predictor explains at most 7.5 % of the differences between participants; the rest stays with the participant.

An angry forward walk of p09 on a 113 cm child body (p13), on the source body and on a 195 cm grid body. Forward lean follows hip width (−6.74°, −3.36°, +3.45°; slope 149.7° per metre against the 150° target), while cadence (2.00 steps/s) and leg-normalised step length (0.69) are identical on all three bodies.

2.2 Optimisation

Body model. Each target body is represented by its SMPL-X skeleton and its rest-pose mesh, partitioned by the skinning weights into fifteen rigid blocks (pelvis, upper torso, head, thighs, shanks, feet, upper arms, forearms and hands), each with a signed-distance field, a mass and a centre of mass. Foot vertices define the sole height, four support corners and the foot length.

Initialisation. The optimisation starts from a geometric copy: joint rotations are copied from the source, the root path is scaled by the ratio of leg lengths and the root height by the ratio of hip heights, and the lowest toe is placed on the floor. The copy preserves cadence and leg-normalised step length, which is what a fixed locomotion command means across bodies.

Geometric copy (left) and optimised retarget (right) of the same clip on a 185 cm, 99 kg body. The copy places the right hand 9 cm inside the torso (maximum penetration 8.9 cm); the optimisation reduces it to 2.6 cm and adapts the forward lean to the body.

Optimisation. Starting from the copy (q̄, r̄), the retargeter solves for joint-rotation corrections δ (axis-angle, per frame and joint) and root offsets ρ,

$$q_{t,j}={\bar{q}}_{t,j}\otimes \exp ({\delta }_{t,j}),r_t={\bar{r}}_t+{\rho }_t,\underset{\delta ,\rho }{\min }E=\sum _k{w}_k{E}_k(\delta ,\rho ),q\_\{t,j\}\; =\; \backslash bar\; q\_\{t,j\}\; \backslash otimes\; \backslash exp(\backslash delta\_\{t,j\}),\; \backslash qquad\; r\_t\; =\; \backslash bar\; r\_t\; +\; \backslash rho\_t,\; \backslash qquad\; \backslash min\_\{\backslash delta,\backslash ,\backslash rho\}\backslash ;\; E\; =\; \backslash sum\_k\; w\_k\backslash ,\; E\_k(\backslash delta,\; \backslash rho),$$

where δ and ρ are uniform cubic B-splines with a control point every three frames (C²-continuous corrections), and forward kinematics on the target body gives joint positions, block frames and foot points. Each term is a mean over the frames of the clip in centimetres, degrees or radians, so that one set of weights serves every clip and every body:

Term	Penalises	Form	Weight
fidelity	departure from the copy	‖δ‖² (rad²) and ‖ρ‖² (cm², vertical × 0.3)	1.0 / 0.01
penetration	body blocks overlapping beyond the allowance	Σ max(0, d − a)² over block pairs and query points; d = depth inside the other block's SDF (cm)	1.0
ground	feet below the floor	Σ max(0, −h)² over the four sole corners, h = height (cm)	1.0
stance	contact height of a planted foot	(min h)² on contact frames	1.0
slide	sliding during contact	‖Δp_foot‖² (cm per frame) on consecutive contact frames	5.0
anchor	drift from the copy's contact anchors	‖p_foot − p̄_foot‖² (cm) on contact frames	10.0
balance	zero-moment point outside the support	max(0, d_ZMP − foot length / 2 − 2 cm)², quasi-static ZMP from the centre of mass	0.3
limits	motion outside the source's joint range	max(0, θ_lo − θ)² + max(0, θ − θ_hi)², source range widened by 0.15 rad	10.0
smooth	non-smooth positions and corrections	‖Δ²p‖² (cm) and ‖Δ²δ‖² (deg), second differences	0.05 / 0.02
effort	acceleration effort over the source's budget	max(0, Ê / Ê_src − 1.25)², Ê = mass-weighted squared block accelerations	1.0
lean	trunk lean not following the body	(θ_f − θ_f*)² + (θ_s − θ_s*)² on the mean forward / side lean (deg)	0.3

The penetration allowance a is the largest of 1 cm, the 95th-percentile penetration of the same block pair on the source body, and the rest-pose overlap on the target body. The lean targets θ_f* and θ_s* shift the copy's mean forward and side trunk lean by k_f·Δw and k_s·Δw for the hip-width difference Δw between target and source (Section 2.1).

Weighted residual of each term over the 400 iterations of one optimisation (a neutral walk of p02 retargeted onto a 175 cm, 100 kg grid body), with the quality checks on the right.

The problem is solved with Adam for 400 iterations. The objectives are soft; physical feasibility is not guaranteed. Every clip is frame-aligned with its source and keeps the persona and context labels of the source.

Outcome. 321,508 clips are optimised. For 4,882 clips the optimisation failed or ended with more penetration than the copy, also after a retry with half the step size and twice the iterations; the geometric copy is stored for these. The 2,570 clips on each participant's own body are the source motion itself.

3. Hands

Representation	Finger joints	Finger motion
smpl_skeleton/smpl	yes (pose_hand)	no: one static hand pose per clip
smpl_skeleton/bvh	yes	no: identical to smpl_skeleton/smpl

The retargeting optimises the body and jaw; the finger rotations are taken from the source SMPL-X fit, which in the current MotionPersona release holds one static hand pose per clip (see the MotionPersona card). The hand meshes are nevertheless part of the body model above, so hand–body and hand–hand penetration is accounted for during retargeting. MotionPersonaX will be updated together with the articulated-hand SMPL-X fits of MotionPersona.

4. Data organisation

MotionPersonaX/
├── manifest.csv                            one row per clip (328,960)
├── bodies.csv                              one row per target body (128)
├── smpl_skeleton/
│   ├── smpl/<body>.tar.zst                 SMPL-X parameters of all 2,570 clips on <body>
│   ├── bvh/<body>.tar.zst                  the same clips as BVH on the SMPL-X skeleton of <body>
│   └── shapes/<body>.npz                   SMPL-X shape coefficients of <body>
└── viewer/                                 visualisation copy for the online viewer (see below)

Because of Hugging Face's limits on the number of files per repository, the 328,960 clips are distributed as one compressed archive per body and representation. The viewer/ folder is a separate, compact copy used only by the online viewer, which streams single clips from it: one file per source clip holding the 23 body joints of all 128 bodies (rotations quantised to int16, about 0.05° error), the pelvis trajectory and the static hand pose. It is not the release format; use the archives in smpl_skeleton/ for the complete SMPL-X parameters and BVH files.

Each archive unpacks to <body>/<clip>.npz (or .bvh), where <clip> is the MotionPersona clip name of the source, e.g. grid_h05_g03/p02_angry_fw.npz is clip p02_angry_fw retargeted onto body grid_h05_g03. Single bodies can be downloaded and unpacked independently:

hf download myshi/MotionPersonaX smpl_skeleton/smpl/grid_h05_g03.tar.zst --repo-type dataset --local-dir .
tar --zstd -xf smpl_skeleton/smpl/grid_h05_g03.tar.zst

4.1 SMPL-X parameters

smpl/*.npz uses the same keys and conventions as MotionPersona (AMASS layout, Z-up, metres, gender-neutral model, flat hand mean): poses (T × 165) = root_orient (3) + pose_body (63) + pose_jaw (3) + pose_eye (6) + pose_hand (90), plus trans, betas (of the target body), gender and mocap_frame_rate. Arrays are stored in float32, the precision of the retargeting output. Eye rotations are zero.

4.2 BVH

bvh/*.bvh contains the 55 SMPL-X joints of the target body in a Y-up frame with centimetre units. The root channels hold the pelvis position in world coordinates and the root offset is zero, so the BVH and the npz describe the same pelvis trajectory and the same frames.

4.3 First frame and alignment with MotionPersona

As in the MotionPersona SMPL-X BVH files, frame 0 of almost every clip (99.7 %) is a T-pose inherited from the retargeting input; it is present in both the npz and the BVH and should be discarded when only the motion is needed. The clips are the MotionPersona SMPL-X motion sampled every second frame: frame k ≥ 1 of a MotionPersonaX clip corresponds to frame 2k − 1 of the MotionPersona SMPL-X npz of the same source clip.

4.4 Frame rate

Clips are sampled at 30 Hz, except the retargeted versions of the 120 MotionPersona clips whose stored frame rate was corrected in that release: 10,240 clips at 60 Hz and 5,120 clips at 15 Hz. Always read the frame rate from the file (mocap_frame_rate, or Frame Time in the BVH).

4.5 Tables

manifest.csv

Column	Description
body, clip	target body and source clip, as in the archive paths
source_subject	participant who performed the source clip
style, style_category, traj	context, context category and locomotion code of the source (see MotionPersona)
frames, fps, duration_s	clip length
status	source (participant's own body), optimised, or copy (geometric copy kept, see Section 2)
pen_max	maximum penetration depth beyond the allowance, cm
effort_ratio	effort proxy relative to the source
bal_viol	fraction of contact frames with the zero-moment-point proxy outside the support region
toe_jitter	mean second difference of the toe during stance, cm per frame²
rot_dev_mean, rot_dev_max	mean and maximum deviation of joint rotations from the geometric copy, degrees
root_dev	mean root displacement from the geometric copy, cm
lean_fwd, lean_fwd_tgt	achieved and targeted mean forward trunk lean, degrees

The quality columns are given for optimised clips only.

bodies.csv: body, kind (captured or grid), height_m, mass_kg, leg_m, hip_width_m, hip_height_m and the ten shape coefficients beta0–beta9.

5. Usage

import numpy as np, smplx, torch

d = np.load('grid_h05_g03/p02_angry_fw.npz')
fps, T = int(d['mocap_frame_rate']), len(d['trans'])
model = smplx.create('path/to/models', model_type='smplx', gender='neutral', num_betas=10,
                     use_pca=False, flat_hand_mean=True)
t = lambda k: torch.tensor(d[k], dtype=torch.float32)
out = model(betas=t('betas')[None].expand(T, -1), global_orient=t('root_orient'), body_pose=t('pose_body'),
            jaw_pose=t('pose_jaw'), leye_pose=t('pose_eye')[:, :3], reye_pose=t('pose_eye')[:, 3:],
            left_hand_pose=t('pose_hand')[:, :45], right_hand_pose=t('pose_hand')[:, 45:],
            expression=torch.zeros(T, 10), transl=t('trans'))

The SMPL-X model files are not distributed with this dataset and must be obtained from the SMPL-X website under their own license.

6. Release processing and limitations

• Verification. For two clips of every body, joint positions from the SMPL-X forward pass of the npz, forward kinematics of the BVH and the retargeting output agree to within 0.01 mm.
• Feasibility. The retargeting objectives are soft constraints; residual penetration, foot sliding and balance violations remain on some clips and are reported in manifest.csv.
• Anonymisation. Identifiers follow MotionPersona; see its card for the anonymisation and consent statements.

License & Commercial Use

This dataset is released under the CC-BY-NC 4.0 License for non-commercial and research purposes only.

If you wish to use this dataset for commercial purposes or enterprise applications, please contact myshi@cs.hku.hk and taku@hku.hk to obtain a commercial license.

Citation

@article{shi2025motionpersona,
  title={MotionPersona: Characteristics-aware Locomotion Control},
  author={Shi, Mingyi and Liu, Wei and Mei, Jidong and Tse, Wangpok and Chen, Rui and Chen, Xuelin and Komura, Taku},
  journal={arXiv preprint arXiv:2506.00173},
  year={2025}
}