rtk-training/bjj-kimura-lesson001

BJJ Kimura from Side Control — Lesson 001

Multi-actor 3D motion capture of a Brazilian Jiu-Jitsu Kimura submission technique from side control, demonstrated by a former IBJJF World Champion (anonymized) with a training partner. Captured at 60 Hz from a calibrated 12-camera mocap rig, 1,215 frames (~20 seconds).

This is the research-tier release: kinematic-only, free, CC BY-NC-SA 4.0. Animation-ready formats (BVH, FBX, mesh geometry) and robot-ready retargeted formats (Unitree G1 NPZ) are available under separate commercial licenses — see License.

What's in this dataset

A single frames.json file containing 3D world-space joint positions for two subjects per frame.

[
  {                                      # frame 0
    "subject_1": {
      "Hips":       [x, y, z],           # 3D position in meters, Y-up
      "Spine":      [x, y, z],
      "RightHand":  [x, y, z],
      ...                                # 56 joints total
    },
    "subject_2": { ... same shape ... }
  },
  // ... 1,214 more frames
]

Tensor shape: loads cleanly to (1215, 2, 56, 3) for (frames, subjects, joints, xyz).

Skeleton (56 joints, full-body with full per-finger articulation):

Region	Joints	Count
Spine	Hips → Spine → Spine1..Spine4	6
Head	Neck, Head	2
Each arm	Shoulder → Arm → ForeArm → Hand → 5 fingers × 3 phalanges	19 × 2 = 38
Each leg	Pelvis → UpLeg → Leg → Foot → ToeBase	5 × 2 = 10

Exact joint names list and bone-length template are in manifest.json.

Quick start

import json
import numpy as np

with open("frames.json") as f:
    frames = json.load(f)

joint_names = list(frames[0]["subject_1"].keys())  # 56 joints
print(f"frames: {len(frames)}, joints: {len(joint_names)}")

# Convert to a (1215, 2, 56, 3) tensor for ML pipelines
data = np.array([
    [[frame[s][j] for j in joint_names] for s in ("subject_1", "subject_2")]
    for frame in frames
])
print(data.shape)  # (1215, 2, 56, 3)

Capture metadata

Field	Value
Technique	Kimura submission from side control (BJJ)
Frame count	1,215
Frame rate	60 Hz (frame_time_sec ≈ 0.01667)
Duration	~20.25 s
Subjects	2 (both subjects perform contact-rich grappling throughout)
Capture rig	12 calibrated cameras, multi-view triangulation
Coordinate system	Right-handed, Y-up, world-space, meters
Origin	Room centroid (mean XZ position over both subjects subtracted)
Actor credential	Former IBJJF World Champion (anonymized — not bound to either subject label)

See manifest.json for the bundle integrity record.

Privacy & anonymization

This dataset is canonically-normalized:

• Every subject across every RTK Motion research-tier dataset shares an identical published skeletal template (canonical bone lengths in manifest.json).
• Verification against this dataset:
  • Within-subject bone-length variance across all 1,215 frames: 0.000000 m (rigid skeleton).
  • Subject_1 vs Subject_2 bone-length difference: 0.00% (identical skeletons across actors).

This means body proportions are not recoverable from the data — the kinematic stream carries motion only, not anatomy. The credentialed actor's identity is not bound to either subject_1 or subject_2 label by data alone (kinematic anonymity preserved by canonical-skeleton normalization; categorical credential preserved at the lesson level as a quality signal).

You can verify this yourself:

import json, math, statistics
with open("frames.json") as f:
    frames = json.load(f)

def dist(a, b):
    return math.sqrt(sum((a[i]-b[i])**2 for i in range(3)))

s1_femur = [dist(f["subject_1"]["LeftUpLeg"], f["subject_1"]["LeftLeg"]) for f in frames]
s2_femur = [dist(f["subject_2"]["LeftUpLeg"], f["subject_2"]["LeftLeg"]) for f in frames]
print(f"subject_1 femur: mean={statistics.mean(s1_femur):.6f}m, std={statistics.stdev(s1_femur):.6f}m")
print(f"subject_2 femur: mean={statistics.mean(s2_femur):.6f}m, std={statistics.stdev(s2_femur):.6f}m")
# Both will print: mean=0.450000m, std=0.000000m

Use cases

• Pose estimation — multi-actor 3D ground truth for two-person interaction tasks
• Action recognition — labeled grappling-technique motion segment
• Sports analytics — biomechanical analysis of submission techniques and contact-rich body interactions
• Robotics imitation learning — kinematic targets for humanoid policies (see commercial-AI tier for retargeted Unitree G1 trajectories)
• Synthetic data augmentation — canonical-skeleton format simplifies retargeting to other body templates

Files in this download

File	Size	Description
frames.json	~10 MB	Per-frame joint positions (the dataset)
manifest.json	~1 KB	Bundle metadata: schema version, frame count, fps, canonical skeleton template, file inventory
LICENSE.txt	~3 KB	CC BY-NC-SA 4.0 license + commercial inquiry path
README.md	this file

License

CC BY-NC-SA 4.0 — non-commercial use, share-alike, attribution required.

For commercial use, the same data is available under paid commercial licenses:

Use case	Channel	Status
Single commercial product, no AI training, no ShareAlike copyleft	Gumroad — rtkmotion.gumroad.com/l/kimura ($299 single-project)	✅ live
Commercial + animator-format bundle (BVH + FBX + mesh)	Fab marketplace listing	pending
AI / ML training rights	Direct license — hello@rtkmotion.io	by request
Per-file agentic access (USDC payments)	rtk-api MCP — connect to https://api.rtkmotion.io/mcp	✅ live
Bespoke captures to your spec	Custom commission — hello@rtkmotion.io	by request

See LICENSE.txt for full CC terms or rtkmotion.io/legal#licensing for commercial license details.

Citation

@dataset{rtkmotion_bjj_kimura_lesson001_2026,
  title     = {BJJ Kimura from Side Control --- Lesson 001},
  author    = {{RTK Motion Intelligence Platform}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/rtk-training/bjj-kimura-lesson001},
  license   = {CC BY-NC-SA 4.0},
  note      = {Multi-actor 3D motion capture, canonical-skeleton normalized}
}

Provenance

Captured 2026 by RTK Motion Intelligence Platform. Anonymized via canonical-skeleton normalization (bvh-data-anonymization-v1) before publication.

A future v1 release will include cryptographic provenance signing (signed bundle manifest + verifiable origin attestation). This v0 release uses the CC BY-NC-SA 4.0 license + the canonical-normalization claim above as the trust anchor. Bundle integrity facts (schema version, frame count, file inventory, canonical skeleton template) are recorded in manifest.json for reference.

Related datasets and channels

This is the first published dataset in the RTK Motion BJJ research-tier collection. Live channels for this same Kimura lesson:

• Research tier (this dataset) — Hugging Face, free, CC BY-NC-SA 4.0 ✅
• Commercial single-project tier — Gumroad, $299: rtkmotion.gumroad.com/l/kimura ✅
• Per-file agentic access — rtk-api MCP: https://api.rtkmotion.io/mcp ✅

Forthcoming for this and future lessons:

• Additional De La Riva and Side Control technique lessons on Hugging Face (research tier, free)
• Commercial animation tier (BVH + FBX + mesh) — Fab listing (pending)
• Commercial-AI tier (retargeted Unitree G1 + future SMPL-X / NVIDIA SOMA) — direct license / Protege.ai data partnership

Browse the live agent-to-agent catalog at api.rtkmotion.io/catalog.