Soft robots can be much more flexible than their rigid parents, but their wide ranges of motion make it difficult to map the location of their body parts.
A new algorithm could give them better control of their movements by optimizing the arrangement of sensors on their body.
“You cannot install an infinite number of sensors on the robot itself,” said Andrew Spielberg, co-lead author of the study, a doctoral student at MIT CSAIL. “So the question is, how many sensors do you have and where do you put them to get the most out of them?”
Researchers have developed a new neural network architecture to answer the question. It works by determining the ideal location for sensors and learning what movements are required for different tasks.
[Read: ‘Sensorized’ skin helps robots understand where the hell they are]
The team first divided the robot’s body into different regions and used their rate of strain as inputs to the neural network. Over time, the network has developed the most efficient sequence of movements to perform tasks, such as which movements to use to enter different options.
The network also tracks the most frequently used regions, then removes the less used areas from the entries in subsequent trials.
This process allows the system to recommend where to place the sensors on the robot.
The researchers tested his suggestions by comparing them to sensor designs made by human experts.
They first asked the roboticists to choose the position of the sensor that they thought would work best for tasks such as picking up objects. Then, they performed simulations comparing the performance of human designs to robots analyzed by algorithm.
Study co-lead author Alexander Amini said he was surprised by the results:
Our model far outperformed humans for every task, although I looked at some of the robot’s bodies and felt very confident where the sensors were going. Turns out there are a lot more intricacies to this problem than we initially expected.
Researchers believe the algorithm could help automate the design of flexible robots and ultimately help machines interact with their environment.
Published March 22, 2021 – 19:28 UTC