Soft robot development could benefit from an algorithm that optimizes sensor placement allowing such machines to better ‘understand’ their environments.
Robots have found themselves adapted to a wide range of tasks ranging from the simple to the complex. Yet, there are still some tasks that even the modern age of robotics still struggles with — tasks involving a more gentle touch than rigid, metallic machines can provide.
The answer to tackling jobs that require extra malleability is soft robots. Such machines — often modeled on living organisms — can also be of great benefit when manipulating irregular objects, working in complex environments or when close to humans. To reliably perform programmed tasks, it is of great use for a soft robot to have an ‘awareness’ of the location of its various parts.
Whilst that might sound relatively straightforward, designing an algorithm that can do this for a machine that can deform its parts in an almost endless variety of ways, is tricky.
Now, researchers from the Massachusetts Institute of Technology (MIT) may have hit upon a solution to this problem. The team, including Ph.D. students in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), Alexander Amini and Andrew Spielberg, have developed an algorithm to assist engineers with sensor placement to optimize soft robotics designs.
They hope that their deep-learning algorithm could, by taking a robot’s design and suggesting the ideal locations for the placement of sensors, help create robots that have a better understanding of the environment in which they function.
The system not only learns a given task but also how to best design the robot to solve that task. Sensor placement is a very difficult problem to solve. So, having this solution is extremely exciting.
Alexander Amini, Ph.D. student, the Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT
The research is due to be discussed at the IEEE International Conference on Soft Robotics and in a paper¹ that will be published in the journal IEEE Robotics and Automation Letters.
Co-Learning of Task and Sensor Placement for Soft Robotics (Teaser)
Soft Robots Without Added Extras
It’s the flexibility of soft robots that ultimately makes them so difficult to model. Soft robot’s rigid metallic counterparts only have so many parts that can move in a limited number of ways resulting in a manageable number of calculations for algorithms that map motion and plan for it.
“The main problem with soft robots is that they are infinitely dimensional,” explains Spielberg. “Any point on a soft-bodied robot can, in theory, deform in any way possible.”
Thus, the number of calculations needed to account for this range of movement rises exponentially, making mapping the position of various parts a real challenge. One way of accounting for this would be upping the number of sensors, but as well as being cost-prohibitive, a near-infinite range of movement means a near-infinite number of sensors.
You can’t put an infinite number of sensors on the robot itself. So, the question is: How many sensors do you have, and where do you put those sensors in order to get the most bang for your buck?
Andrew Spielberg, Ph.D. student, the Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT
Solutions do exist to this problem, but they rely on external measures like video cameras that feed information back to a soft robot’s control system. The MIT team wanted to make a soft robot that is self-sufficient and doesn’t need any external extras to assist it. And they say deep-learning finally makes this achievable.
Optimizing Sensor Placement With a Novel Neural Network
The MIT researchers set about designing a novel neural network to provide optimal sensor placement information, and their initial step was to partition the robot body being tested into different regions — which they termed particles.
These particles individually record data about the strain they are experiencing whilst performing tasks and feed it back to the neural network, which can then use trial and error to learn the most efficient range of movements to complete a specific task.
This learning process extended to discovering which particles were utilized less during the performance of specific tasks, allowing these regions to be omitted from future tests. This process of ranking the most important particles allows the network to tell developers where sensors should be placed on the soft machine.
One exciting result that emerged from the MIT team’s testing was the fact that when their algorithm was pitted against experts in robotics, it could place sensors significantly more effectively than human intuition alone could manage.
“Our model vastly outperformed humans for each task, even though I looked at some of the robot bodies and felt very confident on where the sensors should go,” Amini points out. “It turns out there are a lot more subtleties in this problem than we initially expected.”
The next step for the MIT researchers is planning how to account for the way sensors on robots interact with each other and create a complex interplay.
According to Spielberg, the algorithm could assist in the automation of robot design. A step that could help revolutionize the creation of custom-built robots designed for very specific tasks.
Additionally, the research could help design further algorithms that can assist the movement of both soft robots and their more rigid counterparts.
1. Spielberg. A., Amini. A., Chin. L., et al, , ‘Co-Learning of Task and Sensor Placement for Soft Robotics,’ IEEE Robotics and Automation Letters, [https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9345345]