Data-driven Haptic Modeling of Plastic Flow via Inverse Reinforcement Learning

Most of the natural and man-made objects that we daily encounter undergo permanent or plastic deformation when the applied force exceeds a certain limit. Furthermore, we frequently use this material property to shape objects to desired forms. Current haptic-enabled VR simulators, however, are limited to interaction with elastic objects. In this paper, we aim to provide a real like haptic experience of manipulation with virtual plastic objects. An end-to-end framework is developed enabling the user to collect the haptic feedback data from real material, building the model, and render it in real-time FEM simulation. We model the plastic flow as a closed-loop block-box controller, which is optimized through inverse reinforcement learning trying to mimic the real deformation. The neural network-based controller in our model allows modeling plasticity with arbitrary complexity. To evaluate the proposed system, data from three real materials with various properties were captured and tested. The experimental results revealed the force feedback to be at the reasonable level of realism having relative error below than the human just noticeable difference of force perception.