Kyo Kutsuzawa ^* Minami Matsumoto Dai Owaki Mitsuhiro Hayashibe

• Tohoku University, Sendai, Japan

When humans grasp an object, they can recognize the properties of the object, such as stiffness and shape, through hand sensation. They can also estimate their confidence in the estimated object properties. In this study, we developed a method for multi-fingered hands to estimate both physical and geometric properties, i.e., the stiffness and shape of the object, as well as their confidence levels using proprioceptive signals, i.e. joint angles and velocities. We designed a learning framework based on probabilistic inference that does not need hyper-parameters to balance between estimation of different types of properties. Using this framework, we implemented recurrent neural networks that estimates the stiffness and shape of grasped objects with their uncertainty in real time. We also demonstrated that the trained neural networks can represent the confidence level of estimation, including uncertainty and task difficulty, as variance and entropy. We believe that this approach will benefit reliable state estimation, flexible object manipulation, and combination with probabilistic inference-based decision making.