Robots, liquids and inference : robotic manipulation and estimation of liquids via approximate simulation

Abstract

Liquids can be found almost everywhere. They are typically observed in human environments inside containers whose shape, form and size influence how they should be manipulated and how they will behave. Despite their complex dynamics and the wide range of possible objects with which they can interact, people can still instinctively comprehend how a liquid would act in a particular circumstance and even get insights into its characteristics through simple exploratory interactions. On the other hand, robots have traditionally been designed to perform repetitive tasks under predefined conditions, with minimum intuitive or adaptive capabilities. Naturally, the following question arises: What would it take for robots to adapt as effectively as humans do, particularly in complex scenarios where liquids are involved? This thesis investigates this question by employing approximate simulation mod els to learn manipulation strategies that can be transferred between simulated and real-world contexts and used to comprehend the relevant features of such tech niques. The investigation begins by examining how a robot’s pouring strategy can be adapted to the geometry of the containers by leveraging coarse models of liquid behaviour to reduce spillage. The results suggest that coarse models of simulation are helpful in some cases. Directly learning from the real world seems to work best for simple cups. However, strategies learned in simulation significantly outperform other solutions for challenging containers. The thesis then analyses whether simple probing tactics such as stirring yield relevant estimates about the properties of a liquid and analyses if such estimates help improve the pouring results. This study considers real and simulated liquids with viscosities ranging from water to gel. This investigation shows that stirring estimates are indeed beneficial for spillage reduction. However, it also shows that, due to unmodeled adhesive behaviour, the reality gap increases with liquid viscosity, resulting in inaccurate observations and increased spillage. The final study focuses on understanding the interplay between initial conditions and parameters of a pouring policy. In this case, simulation-based models are used to learn a distribution over the set of conditions or environmental parameters under which the pouring policy is most likely to be effective. Evidence shows that the proposed method can capture complex relations like multi-modality and sharp edges in moderate dimensional spaces.