DirtyMoCap recovers full-body mocap from noisy markers, CUDA solver up to 100x faster
petewoodbridge · x · 2026-10-13
DirtyMoCap (SIGGRAPH Asia 2026, Zhejiang University team) recovers clean 4D SMPL-H motion from sparse, shuffled, noisy optical mocap markers using proxy anchors and a differentiable learnable Gauss–Newton solver.
- One model generalizes across marker configurations; no layout or labels needed
- Handles dropouts, ghost points, jitter; recovers body, hands and shape
- CUDA solver up to 100x faster than PyTorch
- Rebuilt 3 hours of old kung fu mocap; code releasing on GitHub
Related event: DirtyMoCap cleans up noisy motion capture data with 100x faster solver(2 posts)→
More from Embodied
- Warehouse robotics firm Ultra raises $50M Series A after packing 500K+ orders — lukas_m_ziegler · 2026-10-13
- microagi hits $100M annualized run rate 6 months after launching robot-data platform Shift — lukas_m_ziegler · 2026-10-13
- XPENG to sell Iron humanoid in H2 2027, in-house Turing chip claimed to match NVIDIA Thor — lukas_m_ziegler · 2026-10-13
- A Massive Week in Robotics: XPENG, Boston Dynamics and More — lukas_m_ziegler · 2026-10-13
- Dev builds Felix, a detective game living on an M5Stack device with Codex and receipt printer clues — OpenAIDevs · 2026-10-13
- Two gaps in sequence: Koltun's is representation, Malik's is information — contact isn't on the screen — le_james94 · 2026-10-13