Creating Fluid-Interactive Virtual Agents by an Efficient Simulator with Local-domain Control

In the realm of digital twin systems, establishing simulation environments for creating and testing virtual agents has garnered substantial attention across various applications. The obtained control policies endow virtual agents with more realistic behaviors and interactive capabilities, finding applications in both computer animation and robotic control. While rigid-body simulators are widely used for virtual agents, achieving similar feats in fluid environments presents formidable challenges due to high complexity and exorbitant costs. One major reason is that most fluid simulators feature a fixed domain, which struggles to enable agents to freely navigate in an unbounded, obstacle-filled space, especially when computational resources are limited, thus restricting their wide utility for creating virtual agents. In this paper, we introduce a novel fluid-solid interaction simulator grounded in an efficient lattice Boltzmann solver. A key feature of this simulator is a dynamically moving local domain that encircles the agent, offering greater flexibility for obtaining control policy while maintaining efficiency in simulation. Previous methods, which anchored a square moving local domain along with the agent, suffered from severe spurious flows when the agent underwent rapid acceleration especially when the domain had to rotate, such as during a U-turn. This led to inaccurate results and instability. Conversely, we propose a novel domain-tracking method that harnesses optimal control techniques to address this issue. Our approach not only bolsters local-domain simulation stability, but also improves efficiency by employing a slender domain, which broadens the application scope of direct fluid-solid interactions for virtual agents. We validate our method by comparing simulations to physical phenomena and obtaining control policies for various virtual agents to accomplish challenging tasks. This effort culminates in a series of animations that vividly demonstrate the efficacy of the entire framework potentially used in both computer animation and robotics.