Exploring the Environment: Haptic regression
The HR-Recycler project aims at creating a hybrid collaboration environment for disassembly of Waste of Electrical and Electronic Equipment (WEEE) materials, where humans and robots collaborate with the goal of disassembling WEEE products. Even for a single device type (e.g., microwave), recycling plants are interested in many variations that come in different shapes, weights, and disassembly steps. The complexity is further exacerbated when the mechanical structures of these devices are damaged which is often the case for disposed WEEE materials. When disassembling devices without knowing their construction plans, fixtures are easily not detected by visual sensors – either because they are not visible, such as snap-in fixtures, or because they are not properly recognized, such as recessed screws. In these situations, a solution is required for understanding the object and identifying its components. Hence, a major challenge in disassembling an unknown object is to overcome these various sources of uncertainty. The shape of an object can be approximated well from vision and depth sensors. However, the information may be incomplete as visually occluded structures and the material composition that affects mechanical characteristics remains out of reach for such sensors. Thus, we are working at providing a robot with the ability to exploit its force sensing capability to refine the knowledge about these objects by interacting with the object.
Technical Challenges
By observing how robot force applied to the object affects interaction dynamics between them, we can improve the knowledge about the object properties, particularly material properties and estimation of structural features. This approach known as tactile and haptic exploration can significantly improve the results of visual object identification methods for complex objects with hidden structures.
Robotic Solution: Haptic SLAM
Haptic SLAM is dedicated to modeling the environment using haptic sensory data without visual feedback, for which haptic regression is an important technology. The entire space of the environment is divided into grids on which the probabilistic environmental model is constructed. Specifically, the probabilistic density functions of the objects in the environment are updated based on collected haptic data using the Bayesian inference and haptic regression. To improve the efficiency of large state space, occupancy and inference grids are applied and adopted with the discretized representation of the workspace. The represented model is built from iteratively collected sensor data given an initial belief over the geometry of the object. An imminent challenge lies in the refinement of the estimated geometric shape of the underlying object. The critical point is to infer the shape candidates represented via analytical parameterization of different shapes from the initial sensor data and add this prior knowledge to further measurements. As a result, the overall dataset allows not only to update the geometric decomposition but also the material decomposition of the object. An example of the modeling process of an object can be referred in Figure 1, and the modeling results of the object are illustrated in Figure 2.

