Asian Scientist (May 19, 2022) –An eagle’s majestic glide through the air, a dragonfly’s controlled hover over a pond, a stingray’s graceful swim through the depths of the ocean—such engineering marvels of nature are inspiring modern robotics. The researchers everywhere are creating biomimetic robots that attempt to accurately imitate an animal’s natural movements in a particular environment. As published in IEEE Robotics and Automation Letters, researchers from Singapore University of Technology and Design (SUTD) used a form of deep machine learning on a stingray-like soft robot to teach it more efficient and precise forms of movement, allowing the robot to gracefully swim through water.
Teaching robots complicated movements is not easy, and for soft robots this is even harder. Unlike a traditional mechanical robot whose movements can be easily predicted because of its rigid links, the movement of a soft robot is highly dynamic due to a higher range of mobility and the use of softer materials like silicone. This means that predicting precise movements of such robots is tougher. To approach this issue, Dr Pablo Valdivia y Alvarado and his team at SUTD used Deep Neural Networks (DNNs). Valdivia y Alvarado is an assistant professor at SUTD.
DNNs are a more intricate and complex form of machine learning that mimic the way a brain makes decisions by detecting and recognising a pattern of information from a series of inputs. It then produces a predicted output based on the previously learned data. In this case, the DNN was used to teach a soft stingray-like robot to propel itself in a water tank and determine the most efficient and effective method of moving its soft fins through the water.
Why shape the robot like a stingray? Speaking to Asian Scientist, Valdivia y Alvarado explains that this is due to the “high maneuverability that can be achieved with a relatively simple and streamlined body.” The robot can turn along multiple axes – such as up or down, left or right, forward or backward.
The team conducted the experiments by attaching the soft robot to a 3D-printed clamp. The clamp contained a six-axis force/torque sensor to measure the twisting and subsequent movement of the fins in water. As the soft robot moved its fins, the amount of force and torque measured by the sensor was recorded. This was repeated 10 times to produce 100 force/torque data sets from 100 robotic inputs for the DNN to begin learning.
The DNN was given this data set to learn which force and torque values are best suited for rapid and effective movement. From there it was told to predict the movement of the fins from a new set of force/torque data to see if it can successfully mimic previously learned fin movements.
Results from feeding the new data set were promising. The soft robot was able to accurately mimic a series of inputs that were highly similar to the initial inputs given during the start of the experiment. Also, the robot achieved this in a relatively short amount of time. The researchers hope that this study could be a stepping-stone for developing and training marine exploration vehicles to rapidly adapt to the ever-changing conditions in the ocean.
“Our next steps will be to use these models for real-time closed-loop control of free-swimming soft robots to really understand how effective they are in predicting the complex dynamics involved [underwater],” said Valdivia y Alvarado.
Source: Singapore University of Technology and Design; Illustration: Shelly Liew
The article can be found at Li et al. (2022), DNN-Based Predictive Model for a Batoid-Inspired Soft Robot.