MotionPersona

Locomotion generated by a controller trained on MotionPersona. 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.

MotionPersona is an optical motion-capture dataset of human locomotion designed to study how movement depends on the characteristics of the person who moves and on the context in which they move. It comprises 2,773 clips (34.0 h) from 48 participants aged 5 to 68. Every participant performs the same protocol of nine contexts (five emotional states, one physiological state and three movement variants) crossed with seven locomotion types, so that the effect of the individual can be separated from that of the context. Each participant is described by age, gender, height, weight, an SMPL-X body shape and, for 44 participants, a persona annotation.

The dataset accompanies the paper MotionPersona: Characteristics-aware Locomotion Control (arXiv:2506.00173). A companion dataset, MotionPersonaX, retargets every clip onto 128 body shapes.

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

Participants	48 (44 with persona annotation: 20 female, 24 male; age 5–68 y; height 105–189 cm; weight 16–90 kg)
Clips	2,773 on the SMPL-X skeleton (34.0 h); 2,763 of them also on the Vicon skeleton (33.8 h)
Contexts	emotional: neutral, angry, depressed, happy, fear · physiological: drunk · movement variants: swimming, two-foot jump, big step
Locomotion	forward, sideways and backward walking and running; transitions
Capture	Vicon optical system, 29 cameras, 5 m × 5 m volume, 39 markers, 120 Hz
Representations	SMPL-X parameters and BVH on the SMPL-X skeleton; BVH on the Vicon skeleton

1. Data acquisition

Capture setup. Motion was recorded with a Vicon optical motion-capture system of 29 cameras covering an effective capture volume of 5 m × 5 m, at 120 Hz with sub-millimetre accuracy. Each participant wore 39 reflective markers placed on anatomical landmarks according to the Vicon standard marker set. Every session was preceded by a calibration of the system and a check of the marker placement.

Protocol. The capture follows a fully crossed design of nine contexts and seven locomotion commands, giving 63 context × command cells that every participant was asked to perform: forward, backward and sideways walking; forward, backward and sideways running; and transitions between them. Each take is one continuous recording of a single cell, typically 15–80 s long (median 40 s). Participants interpreted each context in their own way, so the data contain both the effect shared by all participants and the participant-specific interpretation. Because the grid is repeated across participants and nearly complete (most participants cover almost all 63 cells), the contributions of the participant, the context and their interaction can be estimated separately.

Processing. Two representations are derived from the marker data. (i) The Vicon skeleton solve is exported as BVH (vicon_skeleton/). (ii) SMPL-X parameters are fitted with MoSh++ (Mahmood et al., 2019) using gender-specific body models; the pose parameters are then combined with per-participant shape coefficients of the gender-neutral model (smpl_skeleton/tools/1_convert_smpl.py), and the SMPL-X-skeleton BVH files are produced from these parameters (smpl_skeleton/tools/2_convert_valid_bvh.py).

2. Participants and annotation

Participants ordered by height, each walking in the neutral context and labelled with the anonymised ID, age, height and the three persona attributes.

Cohort. The 48 participants span 5 to 68 years, 105 to 189 cm and 16 to 90 kg, and include five children and several participants over sixty. Forty-four of them (24 male, 20 female) consented to persona annotation; the remaining four (b15, b27, b33, b34) contribute motion and body shape without persona attributes.

Demographic and anthropometric attributes. Age, gender, height and weight were collected with a questionnaire before the session. The SMPL-X shape coefficients (10 betas) fitted by MoSh++ define each participant's skeleton and mesh and are provided as a complementary, mesh-based description of the body.

Persona attributes. Each annotated participant chose one keyword on each of three dimensions derived from the interpersonal circumplex (Wiggins, 1979; see keywords.txt): a role or life stage (child, student, professional, artist, athlete, performer, homemaker, retiree), an affiliation or sociability (withdrawn, reserved, moderate, sociable, gregarious) and a dominance or assertiveness (submissive, compliant, moderate, assertive, dominant). The keywords describe the person rather than the motion and are therefore constant across all clips of a participant. A free-text description of the person is provided in addition.

Context categories. Following the paper, the nine contexts are grouped into emotional states (neutral, angry, depressed, happy, fear), a physiological state (drunk) and movement variants (swimming, two-foot jump, big step).

Three example participants of different age and build, each performing the nine contexts. Columns share the context; rows share the participant.

Left: age and height of the annotated participants. Right: number of clips per context and locomotion type.

3. Data organisation

MotionPersona/
├── subjects.csv                       one row per participant
├── manifest.csv                       one row per clip
├── keywords.txt                       persona attribute taxonomy
├── smpl_skeleton/
│   ├── smpl/<subject>/<clip>.npz      SMPL-X parameters
│   ├── bvh/<subject>/<clip>.bvh       the same motion as BVH on the SMPL-X skeleton
│   ├── shapes/<subject>.npz           SMPL-X shape coefficients of each participant
│   └── tools/                         conversion scripts (SMPL-X parameters -> BVH)
└── vicon_skeleton/
    └── bvh/<subject>/<clip>.bvh       Vicon skeleton solve

Raw C3D marker trajectories will be added in a later update.

Clips are named <subject>_<context>_<locomotion>, e.g. p02_angry_fw; each participant additionally has a T-pose clip <subject>_tpose on the SMPL-X skeleton. Locomotion codes:

Code	Locomotion	Code	Locomotion
fw	forward walking	fr	forward running
sw	sidestep walking	sr	sidestep running
bw	backward walking	br	backward running
tr	transitions	trs	transitions, side-facing variant (3 participants)
id	idle	mix	single take mixing several locomotion types (1 participant)

3.1 SMPL-X skeleton

• smpl/*.npz follows the AMASS convention for SMPL-X: poses (T × 165, axis-angle) is the concatenation of root_orient (3), pose_body (63), pose_jaw (3), pose_eye (6) and pose_hand (90); trans (T × 3) is in metres; betas (10), gender (neutral) and mocap_frame_rate complete the file. The coordinate frame is Z-up. The hand pose is relative to the flat hand mean and is constant within each clip in this version (see Section 3.3).
• bvh/*.bvh contains the 55 SMPL-X joints in a Y-up frame with centimetre units and the same frames as the corresponding npz. Frame 0 is an additional grounded T-pose inserted for retargeting and should be discarded when only the motion is needed. The root trajectory of the BVH differs from the pelvis trajectory of the npz by a constant offset equal to the rest-pose pelvis position of the participant (about 35 cm horizontally and below 1 cm vertically); joint rotations are identical.
• Frame rate. 2,653 clips are sampled at 60 Hz, 80 clips at 120 Hz and 40 clips at 30 Hz. The frame rate must therefore always be read from the file (mocap_frame_rate, or Frame Time in the BVH).

3.2 Vicon skeleton

• bvh/*.bvh is the skeleton solved by Vicon from the markers: 59 joints including fingers, with generic joint names (Hips, Spine, LeftUpLeg, ...). 2,728 clips are sampled at 120 Hz and 35 at 60 Hz.
• The original export is Z-up with inch units. For this release it was re-expressed as Y-up with centimetre units through a rigid change of frame applied consistently to every joint offset, root translation and local joint rotation, such that the rest pose is also upright. The conversion is lossless: joint positions obtained by forward kinematics agree with the original export to within 0.005 mm.

3.3 Hand articulation

Representation	Finger joints	Finger motion
vicon_skeleton/bvh	yes	yes, as solved by Vicon
smpl_skeleton/smpl	yes (pose_hand)	no: one static hand pose per clip
smpl_skeleton/bvh	yes	no: identical to smpl_skeleton/smpl

In the current SMPL-X fits the hand articulation was not transferred from the capture, so the hand pose is held constant over each clip. SMPL-X fits with articulated hands will be released in a later update. Until then, use vicon_skeleton/ for finger motion.

3.4 Tables

subjects.csv

Column	Description
subject	p01–p44 (annotated participants); b15, b27, b33, b34 (participants without persona annotation)
gender, age, height_cm, weight_kg	questionnaire data
role, affiliation, dominance	persona attributes (see keywords.txt)
n_clips_smpl, n_clips_vicon	number of motion clips on each skeleton
description	free-text description of the participant
description_variants	short, medium and long rewrites of the description, separated by #

manifest.csv

Column	Description
clip, subject	identifiers as in the file tree
style, style_category	context and its category (emotional, physiological, movement; reference for T-poses)
traj	locomotion code
smpl_frames, smpl_fps, duration_s	length of the SMPL-X clip
vicon_frames, vicon_fps	length of the Vicon clip (empty if not available)

4. Usage

import numpy as np, smplx, torch

d = np.load('smpl_skeleton/smpl/p02/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'))
vertices = out.vertices          # (T, 10475, 3), Z-up, metres

The SMPL-X model files are not distributed with this dataset and must be obtained from the SMPL-X website under their own license. BVH files can be imported directly into Blender, MotionBuilder or Maya.

5. Release processing and limitations

• Anonymisation. Participants are identified only by the codes above, and all file names, headers and tables were checked for personal names. Body shape, anthropometry and motion are nevertheless personal characteristics; any attempt to re-identify participants is prohibited.
• Consent. All participants signed a consent agreement prior to capture.
• Selection. The release contains the takes with a valid SMPL-X fit. A Vicon-skeleton version is provided for all of them except ten clips of p04, for which no Vicon export is available.
• Frame-rate metadata. For 120 clips the frame rate stored with the SMPL-X fit was inconsistent with the recording; it was corrected by comparing root velocity against the Vicon take of the same clip. Motion data were not resampled.
• Temporal alignment. SMPL-X fits occasionally omit a few seconds at the beginning or end of a take; the two skeleton versions of a clip are therefore not frame-aligned in general.
• Coverage. Not every participant has every context–locomotion combination (see the figure in Section 2). Participant b15 contributes 17 clips.

6. Related dataset: MotionPersonaX

MotionPersona captures each participant in exactly one body, their own: the diagonal of the persona × body grid. MotionPersonaX fills all 44 × 128 cells.

In MotionPersona, who performs a motion and the body that carries it are inseparable, since each participant is recorded only in their own body. MotionPersonaX removes this coupling: it retargets the 2,570 clips of the 44 annotated participants onto 128 SMPL-X bodies (the 48 bodies of this dataset and 80 synthetic bodies spanning 1.05–1.95 m in height and eight girth levels), giving 328,960 clips (about 4,200 h) in which persona and body vary independently. The retargeting is an explicit optimisation that preserves the timing of the source and adapts posture, contacts, balance and trunk lean to the target body.

The two datasets share identifiers and clip names: <body>/p02_angry_fw.npz in MotionPersonaX is clip p02_angry_fw of this dataset on body <body>, and the clip on the participant's own body is the source motion itself. MotionPersonaX clips are sampled at half the frame rate of the SMPL-X fits (frame k corresponds to frame 2k − 1 here), use the same SMPL-X conventions, and carry the same static hand pose per clip (Section 3.3).

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}
}

References

• N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, M. J. Black. AMASS: Archive of Motion Capture as Surface Shapes. ICCV 2019.
• G. Pavlakos, V. Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, M. J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image (SMPL-X). CVPR 2019.
• J. S. Wiggins. A psychological taxonomy of trait-descriptive terms: The interpersonal domain. Journal of Personality and Social Psychology, 37(3), 1979.