Teaching underwater stingray robots to swim faster and with greater precision using machine learning

Researchers from the Singapore University of Technology and Design (SUTD) developed a new approach to model the dynamics of underwater stingray-like robots using Machine Learning. This approach can enable more efficient swimming in complex underwater environments by accurately predicting required flapping motions for a set of given propulsive force targets.

Their study ‘DNN-Based Predictive Model for a Batoid-Inspired Soft Robot’, published in IEEE-RAL, will pave the way towards better control of autonomous underwater robots.

Bio-inspired soft robots are unique due to their elegant, natural movements. However, modelling and controlling soft robot bodies underwater are challenging due to their infinite degrees of freedom and complex dynamics.

The research team focussed on developing a suitable Deep Neural Network (DNN) model to predict desired flapping motions to achieve the required locomotion in a rapidly changing environment. Unlike traditional physics-based models, DNN models can provide minimal input-output relationships for the complex dynamics found in soft bodies.

Once the team matched the measured propulsive forces generated during DNN model’s predicted flapping sequence against the DNN model’s target forces, they were confident that DNN would be more suitable to predict and accurately mimic the complex physical properties of soft underwater robots.

The experiments were carried out inside a water tank by afixing the robot to a custom-designed 3D printed clamp connected to a 6-axis load cell. The sensor attached to the clamp was used to measure the forces and torques generated during the flapping of the robot’s fins. Different input signals were tested on the robot, and the collected forces and torques provided minimal input-output relationships for the complex dynamics found in soft bodied robots. In total, 10 experiments were conducted to collect 100 Force/Torque data sets from 100 different robot input sequences.

The new approach simplified the otherwise painstaking modelling process and enabled reliable predictions that could be used for programming underwater robots’ flapping sequences to generate desired propulsive forces.

“Our research team will continue to explore trained DNN models using integrated sensors and autonomous behaviour control of robots in dynamic underwater environments for marine inspection and exploration,” said Assistant Professor Pablo Valdivia y Alvarado, principal investigator from SUTD.