🤖 KrumpRobot: Neural Physical AI Motion Policy for Humanoid Street Dance

KrumpRobot is a neural movement generation policy trained on Modal Cloud from monocular in-the-wild video of Arun's live Krump street dance performance (Fractured Cosmos session). It synthesizes authentic Krump kinetic primitives (STOMP, JAB, ARM SWING, CHEST POP, KILL-OFF) conditioned on 140 BPM music beat timing and ground reaction force targets, retargeted directly onto the Adobe Mixamo Y-Bot rig and Unitree H1 / G1 humanoid robots.

• Interactive 3D WebGL Space: https://huggingface.co/spaces/socialprescribing/KrumpRobot
• Training Platform: Modal Cloud (Debian Slim, 4 vCPUs)
• Model Checksum (SHA-256): d066e7f5c602212a9e4d660b3584ac6f675b6012cfd68849d7b6d6a2bb31d061

---

🦾 Architecture & Kinematic Decoders

KrumpRobot uses a causal recurrent architecture with residual multi-task decoders:

[Joint State (182-dim)] ---> [State Proj (192)] --+
                                                  |---> [Fusion (256)] ---> [2-Layer GRU (256)] ---> [ResMLP]
[Conditioning (9-dim)]  ---> [Cond Proj (64)]  --+                                                         |
  - 5 Primitives (One-Hot)                                                                                  +---> [Pose Head]       (22 Quats = 88-dim)
  - 140 BPM Phase (sin, cos)                                                                                +---> [Velocity Head]   (Joint dq/dt = 88-dim)
  - Energy Climax (0 to 1)                                                                                  +---> [Root Head]       (Hips delta = 3-dim)
  - Target GRF F_z (BW)                                                                                     +---> [Dynamics Head]   (cL, cR, Fz, Jerk = 4-dim)
                                                                                                            +---> [ZMP Stability]   (x_zmp, z_zmp = 2-dim)

Empirical Benchmarks

• Training Time on Modal Cloud: 547.06 seconds (100 epochs)
• Final Multi-Task Loss: 0.00976
• Geodesic Angular Error: 0.0405 rad (2.32°)
• Inference Latency: p50 = 2.696 ms, p99 = 4.069 ms (real-time robotic MPC ready)
• Controlled Joints: 22 Mixamo bones / 28-DoF whole-body humanoid kinematics

---

📦 Model Files & Artifacts

File	Description
krumprobot_weights.pt	PyTorch state dictionary (3.8 MB)
krumprobot_weights.json	WebGL JSON weights for zero-install Three.js execution
krumprobot_generated_motion.npz	2,950-step (98.46s @ 30 fps) neural-generated Krump choreographic trajectory
mixamo_character.glb	Skinned 3D animated Y-Bot character mesh (4.9 MB, full 98.46s continuous animation)
krump_track.mp3	Authentic 140 BPM FL Studio 21 Krump soundtrack (98.46s)
receipt.json	Cryptographic training receipt with SHA-256 seal
sample_inference.py	Standalone Python evaluation script

---

🚀 Quickstart

git clone https://huggingface.co/socialprescribing/KrumpRobot
cd KrumpRobot
pip install torch numpy
python sample_inference.py

Citation

@misc{krumprobot2026,
  author = {Arun Nadarasa and Antigravity Team},
  title = {KrumpRobot: Neural Physical AI Motion Policy for Humanoid Street Dance},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/socialprescribing/KrumpRobot}}
}