2025 —
First author
CycloFormer
A rotation-invariant transformer for wrist-sEMG hand pose estimation, used to study when biosignals complement vision under occlusion and how the modality scales with data.
NeurIPS 2026 (under review)
- Built exact ℤ₁₆ invariance to wristband donning rotation through channel-shared TDS-CNNs, circular RoPE, and permutation-invariant attention pooling.
- Achieved new state of the art on emg2pose: a 4M-parameter model beats the previous SoTA 6M model by meta on every generalization split with 33% fewer parameters; a 48M model widens the Stage split margin by 2.5° / 3.3 mm.
- Constructed the first controlled sEMG-vs-vision comparison under fingertip self-occlusion; vision wins when the hand is visible, while sEMG becomes more reliable once two or more fingertips are occluded.
- Fit a data-scarcity-aware scaling law across 8 model sizes and 5 data fractions, explaining 98.8% of PA-MPJPE variance and projecting that closing the vision gap would require roughly 34× more session-hours than emg2pose.