Research map / Video / Video generation
Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control
Weisheng Xu, Qiwei Wu, Jiaxi Zhang, Jing Tan, Yangfan Li, Yuetong Fang, Jiaqi Xiong, Kai Wu, Rong Ou, Renjing Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| iterate difficulty-escalating data generation with controller feedback (loop0 to loop6) | PHC single-primitive tracker · Test sets (2201 clips) average | success rate (%): 55.9 -> 76.9 | Improved |
| train on iteratively synthesized data instead of AMASS | PHC single-primitive tracker · Test sets (2201 clips) average | success rate (%): 58.3 -> 76.9 | Improved |
| apply the iterative framework to MaskedMimic (loop1) | MaskedMimic (FC*) · Video-Convert | success rate (%): 41.6 -> 62.4 | Improved |
| use the first synthesized loop0 data only | PHC single-primitive tracker · Test sets (2201 clips) average | success rate (%): 58.3 -> 55.9 | Hurt |
| add physics tracking metrics to the observation (ablation) | CLAIMS feedback loop · EMDB loop3 | success rate (%): 35.6 -> 64.4 | Improved |
| add VLM feedback to the observation (ablation) | CLAIMS feedback loop · Four third-party benchmarks, loop3 | success rate (%): EMDB 60.0 -> 64.4; AIST++ 81.9 -> 82.1; Video-Convert 57.2 -> 50.9 | Mixed |
| add the manually curated variable template library (ablation) | CLAIMS prompt generation · Four third-party benchmarks, loop3 | success rate (%): Video-Convert 46.8 -> 50.9; EMDB 53.3 -> 64.4; AIST++ 76.5 -> 82.1 | Mixed |
| use feedback-driven iteration instead of one-shot training with equal data size (ablation) | PHC single-primitive tracker · AIST++ (1320 clips) | success rate (%): 85.3 -> 88.1 | Improved |
| use feedback-driven iteration instead of one-shot training with equal data size (ablation) | PHC single-primitive tracker · EMDB (45 clips) | success rate (%): 73.3 -> 64.4 | Hurt |
| raise data difficulty over loops | CLAIMS datasets · PHC+ tracking on own loops | success rate (%): 75.3 -> 53.6 | Improved |
| use expert variable-library prompts instead of random Gemini prompts for MDM (ablation) | MDM synthesis · martial arts t-SNE analysis | overlap with professional motion manifold | Improved |
Closest papers by meaning
- Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.88)
- KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills (NeurIPS 2025 · similarity 0.88)
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.86)
- H-GAP: Humanoid Control with a Generalist Planner (ICLR 2024 · similarity 0.86)
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework (NeurIPS 2026 · similarity 0.85)
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
