Friction from Reflectance: Transfer Learning Approach

Gathering knowledge about the world surrounding the robot is a crucial step towards the robot's autonomy. Part of that knowledge are the physical parameters of the objects, like stiffness, dumping or friction coefficients, which are critical for performing the interaction. Similarly to the human perception system, also for robots, vision is the sense that provides the most data, so one can consider whether it is possible to estimate the parameters mentioned above based on images. In this paper, we are proposing a new approach of estimating friction coefficient from vision, i.e. reflectance images. The solution is based on transfer learning. Understood here as the use of pre-trained networks to solve the friction estimation task. Our results surpass the state-off the art approach on a publicly available dataset. The paper first provides a short overview of the state of the art followed by the description of the dataset. Then, we describe our method and show the obtained results. Finally, the discussion of the results and conclusions are given.