Add Kimodo paper: Scaling Controllable Human Motion Generation

[ImChong]

Summary

Added comprehensive documentation for the Kimodo paper on scalable controllable human motion generation, including detailed analysis of its two-stage diffusion architecture, unified control interface, and 700-hour mocap dataset.

Changes Made

• New paper documentation (papers/14_Human_Motion/Kimodo__Scaling_Controllable_Human_Motion_Generation/Kimodo__Scaling_Controllable_Human_Motion_Generation.md):

  • Complete paper summary with bilingual titles (English/Chinese)
  • Problem statement addressing public mocap data limitations
  • Detailed methodology explanation covering:
    • 700-hour commercial-friendly optical mocap dataset with fine-grained temporal text annotations
    • Two-stage denoising architecture (root/body decoupling) to reduce artifacts
    • Unified, composable control interface supporting text, keyframes, sparse joint constraints, 2D waypoints, and dense paths
    • Multi-skeleton output support (SMPL-X, Unitree G1, SOMA) with post-processing
  • Visual flowchart (mermaid diagram) illustrating the complete pipeline
  • Core contributions and implications for humanoid robot learning
  • Resource references and related reading links

• Metadata updates:

  • Updated _data/papers.json to register Kimodo as paper #520 in 14_Human_Motion module with arXiv 2603.15546
  • Reordered Physics-Based_Animation papers to maintain chronological consistency
  • Updated progress.json to reflect module rotation completion (13 → 14)
  • Added entry to papers/DAILY_SUMMARY_LOG.md documenting completion date (2026-06-13)
  • Updated paper count badges in README.md (610→611 papers, 147→148 notes)

Notable Details

• Paper emphasizes data-first approach: 700 hours of commercial-friendly mocap as primary contribution
• Root/body decoupling in two-stage denoising directly addresses artifact reduction (drift, jitter, foot skate)
• Unified control interface supports flexible constraint composition, relevant for robot teleoperation and multi-command strategies
• Open-source release includes inference code, interactive timeline editor demo, and evaluation benchmarks (Apache-2.0)
• Direct applicability to humanoid robot learning pipelines (GMT, HOVER, SONIC) as reference motion source

https://claude.ai/code/session_019pfouvR8wBNTyAdyJUotKC