KungfuAthleteBot

Dataset Overview

The dataset originates from athletes’ daily martial arts training videos, totaling 197 video clips. Each clip may consist of multiple merged segments. We apply automatic temporal segmentation, resulting in 1,726 sub-clips, ensuring that most segments avoid abrupt transitions that could introduce excessive motion discontinuities.

All sub-clips are processed using GVHMR for motion capture, followed by GMR-based reorientation. After filtering and post-processing, the final dataset contains 848 motion samples, primarily reflecting routine training activities.

Due to substantial noise in the source videos, the dataset has undergone multiple rounds of manual screening and meticulous per-sample refinement. While minor imperfections may still remain in certain samples, we ensure that the vast majority satisfy the requirements for reliable motion tracking.

Project Status: Active Development with Ready Ground Subset Model training is currently under active development. The Ground subset of the dataset is largely complete and ready for training. The Jump subset still has minor imperfections due to video source limitations. Most samples have been carefully screened, though training performance may vary. Your feedback and suggestions are greatly appreciated.

Dataset Category Distribution

Category	Count	Example Subcategories
Daily Training	624	Training(624)
Saber / Sword	111	Saber(50), Sword(37), Southern Saber(15), Tai Chi Sword(9)
Fist	98	Long Fist(69), Tai Chi Fist(21), Southern Fist(8)
Staff	90	Staff Technique(90)
Skills	69	Side Flip(35), Air Spin(18), Lotus Swing(13), Backflip(3)

Category	Total	Subcategory	Count
Ground	822	Daily Training	550
		Fist	93
		Saber / Sword	99
		Staff	80
Jump	170	Daily Training	74
		Fist	5
		Saber / Sword	12
		Skills	69
		- Side Flip	35
		- Air Spin	18
		- Lotus Swing	13
		- Backflip	3
		- Other Skills	0
		Staff	10

Notes

• The dataset consists of 992 motion samples in total, divided into Ground (822) and Jump (170) categories.
• Daily Training accounts for the largest proportion, with 624 samples across both ground and jump motions, mainly reflecting basic routine movements.
• Ground motions include Fist (93), Saber / Sword (99), and Staff (80); jump motions include a small number of Fist (5), Saber / Sword (12), and Staff (10) techniques.
• Skill-based movements only exist in jump motions, totaling 69 samples, focusing on high-difficulty acrobatics such as Side Flip (35), Air Spin (18), Lotus Swing (13), and Backflip (3).
• Weapon-related motions only contain overall movement sequences and do not include detailed weapon or hand operations (to be updated in future versions).

---

Motion Statistics Comparison

All metrics are averaged over the entire dataset.

Dataset	Body Lin Vel(w)	Body Ang Vel(w)	Frames Range
AMASS	0.188±0.165	0.510±0.309	[2,8788]
PHUMA	0.193±0.130	0.585±0.368	[49,222]
LAFAN1	0.790±0.251	1.962±0.553	[3065,9171]
KungFuAthlete (Ground)	0.496±0.324	1.537±0.959	[24,15792]
KungFuAthlete (Jump)	0.948±0.473	2.832±1.431	[25,2842]

---

Ground vs. Jump Subsets

We divide the dataset based on the presence of jumping motions:

• KungFuAthlete (Ground) Contains non-jumping actions, emphasizing:

  • Continuous ground-based power generation
  • Rapid body rotations
  • Weapon manipulation and stance transitions

• KungFuAthlete (Jump) Includes high-dynamic aerial motions such as:

  • Somersaults
  • Cartwheels
  • Other acrobatic jumps

Key Observations

• The Jump subset exhibits the highest joint velocity, body linear velocity, and angular velocity among all compared datasets.
• The Ground subset, while excluding jumps, still shows significantly higher dynamics than natural motion datasets (e.g., LAFAN1, AMASS).
• Compared to PHUMA and AMASS, which focus on daily activities and walking motions, KungFuAthlete demonstrates stronger non-stationarity, larger motion amplitudes, and more challenging transient dynamics, even at comparable or higher frame rates.

⚠️ Safety Warning

The Jump subset is intended for studying the upper limits of humanoid motion. Direct training or deployment on real robots without strict safety constraints may lead to serious hardware damage.

Demo

278 GVMR SMPL

278 GMR, Height-Adjusted

Pipeline

KungfuAthlete's data pipeline:

[Video] → Cut by scene → GVHMR → [GVHMR-Pred (smplh)] → GMR → [Robot Motion (qpos)] → Artificial Selection → Height-Adjusted → [KungfuAthlete]

Supported Robots

Robot	ID	DOF
Unitree G1	unitree_g1	29

See GMR README for other list

Data Format

GVHMR pred

# gvhmr_pred.pt
{
  "smpl_params_global":
  {
    "body_pose": torch.Tensor,    # (N, 63)
    "betas": torch.Tensor,    # (N, 10)
    "global_orient": torch.Tensor,    # (N, 3)
    "transl": torch.Tensor,    # (N, 3)
  }
  "smpl_params_incam":
  ...
}

If you intend to perform motion retargeting to other robotic platforms, or to employ alternative retargeting methodologies, please use the SMPL-H files provided in the collection_gvhmr directory.

GMR qpos (org_smoothed/)

# robot_qpos.npz
{
  "fps": array(30),
  "qpos": np.ndarray,    # (N, 36) 36 = 3(position xyz) + 4(quaternion wxyz) + 29(DOF)
}

For BeyondMimic (org_smoothed_mj/)

All motion files under org_smoothed_mj/ are stored in NumPy .npz format and contain time-sequential kinematic data for Unitree G1 training. Each file represents a single motion clip sampled at fps Hz and includes joint-level states and rigid-body states expressed in the world coordinate frame (quaternion format: xyzw; units: meters, radians, m/s, rad/s).

motion_clip = {
    "fps": np.array([50]),               # Frame rate (Hz)
    "joint_pos": np.ndarray,             # Shape (T, 36), joint DoF positions (rad)
    "joint_vel": np.ndarray,             # Shape (T, 35), joint angular velocities (rad/s)
    "body_pos_w": np.ndarray,            # Shape (T, 53, 3), rigid body positions in world frame (m)
    "body_quat_w": np.ndarray,           # Shape (T, 53, 4), rigid body orientations in world frame (xyzw)
    "body_lin_vel_w": np.ndarray,        # Shape (T, 53, 3), rigid body linear velocities (m/s)
    "body_ang_vel_w": np.ndarray,        # Shape (T, 53, 3), rigid body angular velocities (rad/s)
    "joint_names": np.ndarray,           # Shape (29,), controllable joint names (ordered)
    "body_names": np.ndarray,            # Shape (53,), rigid body names (ordered, includes head_link)
}

Here, T denotes the number of frames in the motion clip.