Figure 2.

General framework for multi-contact information inference: the framework first correlates the 10 real physical strain gauge values with the simulated sensor values using Transfer Net F^T and then predicts the deformation pattern based on the simulated sensor values using Reconstruction Net F^R. According to the predicted deformation pattern, the T multi-contact's candidate locations are extracted using k-Max method. Depending on the rigidness distribution of the 3D structure, the force magnitude of each candidate contact is predicted based on the location and the deformation using Sensitivity Net F^S. All predicted candidate force information is then summarized into robot system for further post-processing to cancel out the spurious contacts.