FiveHundredShots: 500 tennis strokes for the Unitree G1

Jin Cheng · Computational Robotics Lab · ETH Zurich

FiveHundredShots is a set of 500 tennis groundstroke clips for the Unitree G1 humanoid (29 joints) holding a racket in its right hand. Each clip is a few seconds of whole-body reference motion around one ball contact, with the contact frame, the contact point and the hit direction annotated.

The motions come from videos of real players in the CalTennis dataset. They were reconstructed with GVHMR, retargeted to the G1 with GMR, and the right wrist was then re-solved so that the racket face points along the hit direction at contact.

Clips	500 (285 forehand, 215 backhand)
Distinct physical strokes	399 (see same-stroke pairs)
Total length	72,886 frames, 24.3 minutes at 50 Hz
Clip length	71 to 201 frames; 405 clips are 151 frames (3 s) with contact at frame 100
Per frame	281 float32 values: root pose, 29 joint angles, velocities, 30 body poses
Robot	Unitree G1, 29 DoF, ball-shaped hands, racket fixed to the right hand
Racket-head speed at contact	median 9.7 m/s, up to 15.6 m/s
Contact height	0.35 m to 1.25 m, median 0.76 m
Size	about 110 MB

Snapshots

One forehand (shot_0458), from 1.2 s before contact to 0.6 s after:

One backhand (shot_0122), same timing:

Eight different clips at their contact frame (top row forehands, bottom row backhands):

The yellow sphere is the centre of the racket string bed, the red sphere is where it is at contact, and the green arrow is the hit direction.

Contents

FiveHundredShots/
├── README.md
├── LICENSE                  CC BY-NC 4.0
├── metadata.jsonl           one row per shot (this is the table the dataset viewer shows)
├── motions/
│   ├── shot_0000.npy … shot_0499.npy
│   └── index.json           the same rows, keyed by file name
├── robot/
│   ├── g1_29dof_tennis.xml  MuJoCo model: G1 + ball hands + racket
│   ├── scene_tennis.xml     the model on a floor, with lights
│   ├── meshes/
│   └── LICENSE              Unitree BSD-3-Clause
├── scripts/
│   ├── fivehundredshots.py  loader (numpy only)
│   ├── visualize.py         MuJoCo viewer and offscreen recorder
│   └── README.md            visualizer guide
└── assets/                  images used on this page

Quick start

pip install huggingface_hub numpy mujoco "imageio[ffmpeg]"
hf download catachiii/FiveHundredShots --repo-type dataset --local-dir FiveHundredShots
cd FiveHundredShots
python scripts/visualize.py

The viewer is described in scripts/README.md.

Load a clip:

import sys
sys.path.insert(0, "scripts")
from fivehundredshots import load_clip, load_metadata

shots = load_metadata()
shot = shots[0]
clip = load_clip(shot["file"])

clip.joint_pos.shape          # (176, 29)
clip.body_pos.shape           # (176, 30, 3)
shot["stroke"], shot["hit_frame"]    # ('forehand', 125)

Or without the helper: np.load("motions/shot_0000.npy") returns the (T, 281) array described below.

Clip format

Every clip is a float32 array of shape (T, 281), one row per frame at 50 Hz.

Columns	Field	Shape	Notes
0:3	root_pos	(T, 3)	pelvis position, metres
3:7	root_quat	(T, 4)	pelvis orientation, (w, x, y, z)
7:36	joint_pos	(T, 29)	radians, joint order below
36:39	root_lin_vel	(T, 3)	world frame, m/s
39:42	root_ang_vel	(T, 3)	world frame, rad/s
42:71	joint_vel	(T, 29)	rad/s, joint order below
71:281	body poses	(T, 30, 7)	per body: position (3) then quaternion wxyz (4), world frame, body order below

Conventions:

• The world frame is z-up with the floor at z = 0. Its origin and heading differ from clip to clip and are not aligned with the court.
• Velocities are central finite differences of the positions at 50 Hz.
• The body poses are the forward kinematics of the root pose and joint angles of the same frame, so they carry no extra information; they are stored for convenience.
• The racket and the ball hands are not in the body list. Their poses follow from right_wrist_yaw_link through the fixed transforms in the robot model.

Joint order (29)

 0 left_hip_pitch_joint        10 right_knee_joint             20 right_shoulder_yaw_joint
 1 right_hip_pitch_joint       11 left_shoulder_pitch_joint    21 left_elbow_joint
 2 waist_yaw_joint             12 right_shoulder_pitch_joint   22 right_elbow_joint
 3 left_hip_roll_joint         13 left_ankle_pitch_joint       23 left_wrist_roll_joint
 4 right_hip_roll_joint        14 right_ankle_pitch_joint      24 right_wrist_roll_joint
 5 waist_roll_joint            15 left_shoulder_roll_joint     25 left_wrist_pitch_joint
 6 left_hip_yaw_joint          16 right_shoulder_roll_joint    26 right_wrist_pitch_joint
 7 right_hip_yaw_joint         17 left_ankle_roll_joint        27 left_wrist_yaw_joint
 8 waist_pitch_joint           18 right_ankle_roll_joint       28 right_wrist_yaw_joint
 9 left_knee_joint             19 left_shoulder_yaw_joint

This is the breadth-first order that Isaac Lab reports for the G1. MuJoCo orders the joints depth-first, so assign them by name (the loader and the viewer do).

Body order (30)

 0 pelvis                      10 left_knee_link               20 left_shoulder_yaw_link
 1 left_hip_pitch_link         11 right_knee_link              21 right_shoulder_yaw_link
 2 right_hip_pitch_link        12 left_shoulder_pitch_link     22 left_elbow_link
 3 waist_yaw_link              13 right_shoulder_pitch_link    23 right_elbow_link
 4 left_hip_roll_link          14 left_ankle_pitch_link        24 left_wrist_roll_link
 5 right_hip_roll_link         15 right_ankle_pitch_link       25 right_wrist_roll_link
 6 waist_roll_link             16 left_shoulder_roll_link      26 left_wrist_pitch_link
 7 left_hip_yaw_link           17 right_shoulder_roll_link     27 right_wrist_pitch_link
 8 right_hip_yaw_link          18 left_ankle_roll_link         28 left_wrist_yaw_link
 9 torso_link                  19 right_ankle_roll_link        29 right_wrist_yaw_link

Metadata

metadata.jsonl has one JSON object per shot; motions/index.json holds the same objects keyed by file name.

Field	Meaning
id	shot_0000 … shot_0499
file	path of the clip, e.g. motions/shot_0000.npy
stroke	forehand or backhand, labelled by hand
fps	50
n_frames, duration_s	clip length
hit_frame, hit_time_s	frame index (0-based) and time of ball contact
sweetspot_hit_world	position of the string-bed centre at hit_frame, world frame, metres
hit_direction_world	unit vector the racket face points along at hit_frame, world frame
same_stroke_as	id of another clip showing the same physical stroke, or null
source_session, source_video	CalTennis session and video the clip comes from
source_segment	part of that video: seg<run>[c<chunk>]_<start>-<end>, start and end in seconds
source_clip	name of the clip in the processing pipeline

Same-stroke pairs

The 500 clips show 399 different strokes. 101 strokes appear twice, and same_stroke_as links the two clips:

• 100 two-view pairs. The stroke was filmed by two cameras, and each video was reconstructed and retargeted on its own. The two clips are independent estimates of one motion, in different world frames. They differ by 9.8° joint RMS on median (right wrist excluded), which is a direct measure of the reconstruction noise in this dataset.
• 1 same-video pair. shot_0027 is a shorter cut of shot_0028.

Keep both clips of a pair on the same side of any train/test split, or drop one of them.

Robot model

robot/g1_29dof_tennis.xml is the Unitree G1 29-DoF model with three changes:

• Ball hands. Both hands are replaced by a rigid ball end effector (left_ball_hand, right_ball_hand, 0.31 kg each) fixed to the wrist yaw link.
• Racket. A racket body (tennis_racket) is fixed to the right ball hand. The model racket is 0.635 m long, with mass 0.32 kg, balance point 325 mm from the butt and swing inertia 0.0166 kg m². It has a capsule and an ellipsoid collider on contact bit 2, so it collides with objects placed on that bit (a ball) but not with the robot or the floor.
• Sweet-spot site. The site racket_sweetspot marks the centre of the string bed. sweetspot_hit_world is this site's position at hit_frame. The string face normal is the −y axis of the tennis_racket body; hit_direction_world is that normal for forehands and its opposite for backhands.

The joint angles also replay on a stock 29-DoF G1; only the racket-related quantities need this model.

How it was made

1. Source video. CalTennis videos of practice and match play, filmed by fixed cameras. Passages with a single player in view were found with a person detector and cut into chunks of at most 30 s at 30 fps.
2. Human motion. GVHMR reconstructed the player from each chunk in a gravity-aligned world frame. Each video was processed on its own (monocular, static camera); the multi-view calibration of CalTennis was not used.
3. Retargeting. GMR mapped the SMPL-X motion to the G1 at 30 Hz by inverse kinematics.
4. Shots. Candidate contacts were taken at peaks of the retargeted hand speed and confirmed, moved or removed by a person watching the source video next to the robot. Each confirmed contact gave one clip, by default from 2 s before to 1 s after it.
5. Review. Clips were inspected one by one and kept only if the robot motion matched the video; 500 remain. Forehand/backhand labels were assigned by hand.
6. Clean-up. Each clip was smoothed (Savitzky–Golay, 5 frames at 30 Hz), and the root height was shifted over time so that the lower foot rests on the floor.
7. Export. Clips were resampled to 50 Hz (linear, slerp for the root orientation), body poses were computed by forward kinematics and velocities by finite differences.
8. Wrist. Monocular reconstruction does not recover the wrist well enough to orient a racket. For every clip the three right-wrist joints at the contact frame were solved so that the racket face points along the horizontal travel direction of the hand, with the racket held as level as the wrist comfortably allows. This correction is blended into the original wrist motion with a smooth (C2) bump around contact. No other joint is changed.

Known limitations

• Contact is inferred. There is no ball in the data. hit_frame is the confirmed peak of hand speed, and hit_direction_world is the racket-face normal after the wrist solve. It is horizontal by construction and is not a measured ball trajectory.
• The right wrist is solved, not observed. The correction at contact is large (median 45°, up to 101°). Treat the wrist motion as a plausible way to present the racket, not as what the player did.
• Right-handed racket. The racket is always in the robot's right hand, and labels follow that. The left arm is retargeted as reconstructed and is not attached to the racket, also in two-handed backhands.
• Monocular reconstruction. Players are far from the camera. Two reconstructions of the same stroke differ by about 10° joint RMS (see above). The source is 30 fps, so the 50 Hz clips contain no detail above that rate.
• Global trajectory. The reconstructed root height drifted and was corrected by the grounding step (median correction 0.14 m; above 0.30 m in 74 clips). This pins the lower foot to the floor at all times, so jumps and flight phases are flattened. Horizontal motion is left as reconstructed, and feet can slide.
• Kinematic only. The clips are not physically simulated and have not been checked for dynamic feasibility or self-collision.
• Joint limits. Smoothing leaves small overshoots of the model's joint limits in 11% of the frames, at most 0.026 rad (1.5°), almost all in the ankles.
• Overlap between neighbours. Consecutive strokes of a rally are often less than 3 s apart, so the end of one clip can cover the same moments as the start of the next: 92 pairs of clips share a median of 0.28 s (at most 0.92 s), 2% of all frames.
• Imbalance. 287 of the 500 clips come from one recording session (01_23_2026_17_00_court2), and the set covers 7 sessions in total. For a held-out set, split by source_session, or at least by source_video and source_segment while keeping same-stroke pairs together.

Licence

The motion data, metadata, images, and the ball-hand and racket meshes are released under CC BY-NC 4.0: free to use and adapt for non-commercial purposes with attribution. The non-commercial condition follows from the sources: CalTennis is CC BY-NC 4.0 and GVHMR is licensed for research and non-profit use.

• The Unitree G1 model and meshes in robot/ are © Unitree Robotics, BSD-3-Clause (robot/LICENSE).
• The code in scripts/ is MIT (scripts/LICENSE).
• No video, image or body-model parameter of any player is included.

Citation

If you use this dataset, please cite it and the work it builds on.

@misc{cheng2026fivehundredshots,
  title        = {FiveHundredShots: 500 Tennis Strokes for the Unitree G1},
  author       = {Cheng, Jin},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  note         = {Computational Robotics Lab, ETH Zurich}
}

@article{demler2026caltennis,
  title   = {CalTennis: Large Multi-View Tennis Video Dataset and Benchmark of Monocular-to-3D Pose Estimation},
  author  = {Demler, Ilona and Xie, Xinran and Werner, Blake and Szczuka, Anna and Perona, Pietro},
  journal = {arXiv preprint arXiv:2606.20542},
  year    = {2026}
}

@inproceedings{shen2024gvhmr,
  title     = {World-Grounded Human Motion Recovery via Gravity-View Coordinates},
  author    = {Shen, Zehong and Pi, Huaijin and Xia, Yan and Cen, Zhi and Peng, Sida and Hu, Zechen and Bao, Hujun and Hu, Ruizhen and Zhou, Xiaowei},
  booktitle = {SIGGRAPH Asia Conference Proceedings},
  year      = {2024}
}

@article{joao2025gmr,
  title   = {Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking},
  author  = {Araujo, Joao Pedro and Ze, Yanjie and Xu, Pei and Wu, Jiajun and Liu, C. Karen},
  journal = {arXiv preprint arXiv:2510.02252},
  year    = {2025}
}

Contact

Jin Cheng, Computational Robotics Lab, ETH Zurich.