Trading Sensing for Structure: Sparse IMU-EMG Fingertip Force Estimation via Neuro-inspired Structured Modeling

We present a Neuro-inspired Structured Model (NiSM) for fine-grained finger-level contact force estimation during dexterous grasping from sparse wearable sensing. NiSM uses a thumb-mounted inertial measurement unit (IMU) and a single-channel wrist electromyography (EMG) sensor, while deriving grasp context from an IMU-based hand-shape latent. This work studies the trade-off between sensing density and structural inductive bias: when sensor coverage is limited, inferred grasp context, kinematic coupling, and muscle-effort modulation can provide useful constraints for force prediction. Inspired by human sensorimotor control, NiSM organizes computation into complementary pathways: a context-conditioned feedforward pathway maps thumb kinematics to finger-specific estimates through sparse learned gating, while an EMG-conditioned modulation pathway provides effort-dependent scaling. By explicitly encoding these structural priors, NiSM reduces reliance on dense sensor arrays and large generic black-box models. Evaluations on multi-user, multi-object datasets show that NiSM achieves performance competitive with a dense-sensing reference where comparable dense inputs are available, while using substantially fewer parameters, and is more robust than generic models under limited-data settings. These results indicate that structured model design can recover part of the information typically supplied by denser sensing, offering a practical direction for wearable finger force estimation under real-world deployment constraints.