Contact-rich insertion tasks demand precise alignment, but perception, geometry, and control uncertainties often cause unintended contact. Sampling-based Model Predictive Control (SMPC) is attractive here because it optimizes through forward simulation without needing gradients through contact dynamics. However, standard SMPC methods typically propose new actions using generic distributions that do not exploit the very contact information that could reveal where to search next.
ContactKernels, introduced in this preprint, addresses that gap by using geometry-conditioned wrench measurements to estimate the contact location and local contact geometry. The estimated contact state then adapts the proposal distribution: the contact location biases search toward corrective motions, while the contact normal shapes the sampling covariance. This lets the controller use contact not just as feedback, but as a direct guide for where to look for a successful insertion.
The authors evaluated ContactKernels with two SMPC controllers—Predictive Sampling and Model Predictive Path Integral—across multiple insertion geometries in simulation and on a real robot. In simulation, mean insertion success improved by over 50 percentage points for both controllers. On the real robot, success increased from 53.3% to 90.0%, and mean insertion time decreased by 33%. The paper is a preliminary preprint, and the results suggest that contact-aware proposal kernels could be a simple and effective upgrade for contact-rich manipulation pipelines.