Rapid Flow Cup-Enabled Liquid Perception Using a Position-Based Physics Simulator for Robotic Liquid Manipulation

Accurate robotic liquid manipulation has been a challenging task for many industry sectors, which requires the robot to sufficiently understand the liquid flow behaviour. Physics simulation-informed liquid perception and visual or tactile-based direct liquid sensing have been explored as useful approaches for liquid manipulation. However, these approaches may not be practical to implement in some resource and space-limited cases as they require a physical robot to learn the liquid flow behaviour through physical interaction with the liquid. This paper proposes a liquid perception framework where a low-cost and robot-free flow cup test (a standard fluidity characterisation test) is utilised to capture the physical liquid flow behaviour which is then transformed into a position-based physics simulation during a virtual flow cup test using Bayesian optimisation. Such liquid perception through the ’real-to-sim’ flow cup test is intended for guiding robot operations in any subsequent liquid manipulation tasks. The viability of the proposed framework is examined by comparing the physical liquid manipulation performance (i.e., accuracy) with that in simulation using the learned liquid flow behaviour. A robotic crack sealing experiment is implemented as a validation use case in manufacturing. The results suggest that the proposed framework is able to capture and predict a random liquid flow behaviour at a statistical mean accuracy of 85.4-87.0% (with a high-probability accuracy of 88-90%) in approx. 6 minutes using the computation resource in this study, validating its feasibility for underpinning general robotic liquid manipulation applications.