FishGym: A high-performance physics-based simulation framework for underwater robot learning

Inspired by biology Underwater robots have many advantages, including strong maneuverability and efficient propulsion. However, training their behaviors in real-world environments is challenging because of the high material cost. Therefore, a simulation platform for the training control policies of these robots would be very useful for researchers.

Image credit: arXiv: 2206.01683 [cs.RO]

A recent arXiv.org paper proposes a high-performance simulation platform that targets the bidirectional interaction dynamics between a fish-like underwater robot and its surrounding fluid environment. It uses a GPU-accelerated lattice Boltzmann solver for high-performance fluid structure interaction modeling in a local motion frame of reference.

The researchers propose an algorithm specifically designed for aquatic biorobots that can obtain natural and effective control policies when swimming. Finally, a set of benchmarking tasks is provided to evaluate and compare different learning methods and control policies.

Bionic underwater robots have proven their superiority in many applications. However, training their intelligence for a variety of tasks that mimic the behavior of aquatic organisms poses several challenges in practice, mainly due to the lack of large amounts of available training data as well as the lack of available training data. as high cost in actual physical environment. In addition, simulation has been considered as a viable and important tool for data collection in different environments, but it is mainly aimed at hard and soft body systems. There is currently a lack of work on more complex fluid systems that interact with immersed solids that can be efficiently and accurately simulated for robotics training purposes. In this paper, we propose a new platform named “FishGym”, which can be used to train fish-like underwater robots. The framework consists of a robotic fish modeling module using an articulated body with a skin layer, a locally coupled two-dimensional fluid structure interactive simulation module based on on the GPU that handles both finite and infinite domains, as well as a reinforcement learning module. We leveraged existing training methods with the ability to adapt the fish-like robot underwater and obtained learned control policies for many benchmarking tasks. Proven training results with reasonable motion trajectories, with comparisons and analyzes with experimental models as well as known real-life swimming behaviors of fish to highlight the advantages of the platform. offer.

Research articles: Liu, W., Bai, K., He, X., Song, S., Zheng, C., and Liu, X., “FishGym: A High Performance Physics-Based Simulation Framework for Under Robotic Learning country”, 2022 Links: https://arxiv.org/abs/2206.1683

You can provide your link to a page relevant to the topic of this post.