nosuke113/caltennis-g1-hitting-references

CalTennis-G1 Hitting References

Related releases: companion human 3D joints https://huggingface.co/datasets/nosuke113/caltennis-human-stroke-joints · code https://github.com/humanoid-tennis-project/g1-tennis-swing

Kinematic reference motions of tennis strokes for the Unitree G1 humanoid (27 DoF, racket rigidly fixed to the right wrist), retargeted from 3D human motion that was estimated from the public CalTennis videos. Every clip carries one hand-annotated hit time and a racket target (racket-center position, velocity and face normal at that frame).

• 474 clips (forehand 232, backhand 131, volley 83, overhead 28; serves excluded), 50 Hz, 0.94-6.54 s (median 2.38 s), 56,861 frames = 18.95 min per variant.
• Two variants of every clip: original (as retargeted) and foot_locked (stance-foot sliding removed, see below). 948 NPZ files, about 100 MB.
• From 21 CalTennis recordings in 10 sessions.

This is a derived, unofficial dataset. It is not endorsed by the CalTennis authors, Unitree, or the authors of GVHMR, GMR or AdaPT.

Intended use

• Training and evaluating motion-tracking / hitting low-level controllers for a humanoid in physics simulation (it was built for a Stage-1 hitting policy that must reproduce a stroke and bring the racket to an annotated target at the annotated time).
• Motion analysis and retargeting research on tennis strokes.

Out of scope: it contains no ball, so it cannot be used to learn or evaluate ball returning; it is not a measurement of the real players or their rackets; it is not validated on hardware. Do not use it to try to identify the people in the source videos. Non-commercial use only (see License).

Files

metadata.jsonl              one row per (clip, variant); 948 rows
data/original/clip_XXXX.npz
data/foot_locked/clip_XXXX.npz
recordings.json             CalTennis source video for each recording_id (path, sha256, pinned URL)
racket_mount.json           how the racket is attached to right_wrist_yaw_link
foot_lock.json              foot-lock parameters and summary

NPZ arrays (load with np.load(path, allow_pickle=False))

key	shape	meaning
fps	(1,)	50
joint_names	(27,)	joint order (legs 2x6, waist_yaw_joint, arms 2x7 incl. wrists)
joint_pos	(T, 27)	joint angles, rad
joint_vel	(T, 27)	joint velocities, rad/s (computed from the trajectory, not measured)
body_names	(14,)	tracked bodies; index 0 is pelvis (root)
body_pos_w	(T, 14, 3)	world positions, m, Z-up
body_quat_w	(T, 14, 4)	world orientations, quaternion wxyz
body_lin_vel_w	(T, 14, 3)	m/s
body_ang_vel_w	(T, 14, 3)	rad/s

metadata.jsonl fields

clip_id (anonymized, shared with the companion human-joint release), variant / foot_locked, file, sha256, stroke_type, frames, fps, duration_s, hit_frame, hit_time_s, hit_window_frames (hit frame +-2), hit_window_truncated, hit_evidence (visible_contact: contact visible in the video, 326 clips; estimated_swing: hit time estimated from the swing because contact was not visible, 148 clips), hit_uncertainty_s (annotator's stated uncertainty, median 0.10 s, max 0.27 s), hit_source_frame_30hz, racket_center_target_m, racket_center_velocity_target_m_s, racket_face_normal_target (in the same world frame as the NPZ of that variant), recording_id, source_frame_start_30hz, source_frame_end_30hz (clip interval in the source video on a 30 Hz frame grid), human_clip_id, ball_state (not_available), physics_verified (false), and for foot_locked rows foot_lock_offset_xy_at_hit_m.

import json, numpy as np
rows = [json.loads(l) for l in open("metadata.jsonl", encoding="utf-8")]
r = next(r for r in rows if r["foot_locked"] and r["stroke_type"] == "forehand")
with np.load(r["file"], allow_pickle=False) as m:
    q = m["joint_pos"]            # (T, 27)
    root = m["body_pos_w"][:, 0]  # pelvis, (T, 3)
print(r["hit_frame"], r["racket_center_target_m"])

The racket target is the racket center (center_in_wrist_frame_m in racket_mount.json, rigidly attached to right_wrist_yaw_link) evaluated on the retargeted reference at hit_frame. It is not a measured racket or ball position.

How it was built

1. Human 3D motion: GVHMR (rev ee960bb6) on CalTennis video (revision 4af17dd4d398aad4116a35b5a74de4388658d356), SMPL-X neutral body, gravity-aligned world frame. Stroke intervals were cut and visually selected (544 clips).
2. Retargeting: GMR (rev bb1bbe40) to the G1 model with a right-hand racket from AdaPT (rev 615b42f8); resampled to 50 Hz (linear positions, slerp orientations). 519 clips passed visual review.
3. Hit annotation: one annotator marked the racket-ball contact time on the source video for each clip; mapped to the nearest 50 Hz frame. Serves (45) were excluded, leaving 474 rally strokes.
4. Foot lock (foot_locked variant): at each frame the lower ankle is treated as the stance foot if its height is below 0.20 m (95.5% of frames); its horizontal velocity is subtracted from all bodies and integrated, so the stance foot stays planted. Only horizontal body positions / linear velocities and the hit targets change; joint angles, heights and orientations are identical to original. The removed travel at the hit frame is 0.43 m on average (median 0.34 m).

Known issues and limitations

• Video-derived foot sliding. Monocular video estimation slides the support foot (median about 0.30 m/s). A planted robot cannot reproduce that pelvis travel, which made some original targets unreachable. foot_locked is a simple rigid fix, not a contact-aware optimization; it does not enforce physical feasibility and can leave other artefacts.
• No ball. No ball position, velocity or spin. The racket target is a kinematic surrogate.
• Hand-annotated hit times by a single annotator, with no inter-annotator check; 148 clips are estimated_swing. Timing uncertainty up to 0.27 s.
• Not physically verified. Retargeting is kinematic; references may violate dynamics, contain self-contact, or exceed velocity limits. physics_verified is false for all clips.
• Estimation and retargeting error. Far camera (about 13-17 m), occlusion and fast arm motion affect depth, foot contact and wrist pose; the robot's proportions change the form.
• Imbalance. Two recordings contribute 97 and 91 clips; players are not labeled and no player-disjoint split is provided. No official train/test split.
• Racket geometry and attachment are a modeling choice, not the players' real racket.

License and attribution

Released under CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/), following the license of the source videos. Changes were made: 3D pose estimation, retargeting to a robot, resampling, foot-locking, and added hit annotations.

• Source videos: CalTennis by Ilona Demler, Xinran Xie, Blake Werner, Anna Szczuka, Pietro Perona (Caltech), CC BY-NC 4.0, https://huggingface.co/datasets/demalenk/caltennis, paper https://arxiv.org/abs/2606.20542. Please cite CalTennis when using this data.
• Human motion estimated with GVHMR (Zhejiang University; code licensed for education/research/non-profit use) using the SMPL-X body model (Max Planck Institute for Intelligent Systems). No SMPL-X model files, shape coefficients or SMPL-X parameters are included.
• Retargeting with GMR (MIT). Robot/racket model from AdaPT (Apache-2.0). Robot: Unitree G1. Model meshes are not included.

@article{caltennis2026,
  author  = {Demler, Ilona and Xie, Xinran and Werner, Blake and Szczuka, Anna and Perona, Pietro},
  title   = {CalTennis: Large Multi-View Tennis Video Dataset and Benchmark of Monocular-to-3D Pose Estimation},
  journal = {arXiv:2606.20542},
  year    = {2026}
}

---

日本語の概要

CalTennisの公開映像からGVHMRで推定した人体動作を、GMRで右手首にラケットを固定したUnitree G1（27関節）へ変換した参照動作です。サーブを除く474クリップ（フォア232・バック131・ボレー83・スマッシュ28、50Hz、計18.95分）に、人手で付けた打撃時刻と、その時刻のラケット中心の位置・速度・面の向きの目標を付けています。各クリップに、変換したままのoriginalと、支持足の滑りを除いたfoot_lockedの2版があります。

用途はシミュレーション内での動作追従・打撃の下位方策の学習と評価です。ボールの情報はありません。 打撃時刻は1名による手付けで、148件は接触が映像で見えずスイングから推定しています。映像推定に由来する足滑りがあり、foot_lockedはその簡易な補正です。物理的な実行可能性は検証していません。

ライセンスは元映像に合わせてCC BY-NC 4.0（非営利）です。利用時はCalTennisを引用してください。SMPL-Xのモデルファイル・形状係数・パラメータは含みません。