Self-sensing components paves way to functional soft robotics
Material able to self-sense its deformations may lead the way to more adaptable and functional soft robotics. Khai Trung Le talks to Ilse Van Meerbeek about her work.
An elastic silicone foam capable of self-sensing its distortions and deformations through machine-learning has been developed by a team at Cornell University, USA, in an attempt to enhance the functionality of soft robotics. Currently, self-sensing robotics involves mounting or embedding sensors that measure a specific type of deformation and does not measure the degrees of deformation. However, the team believes that embedding self-sensing into the actual material enables superior response to its environment, damage and recovery.
The team, led by Ilse Van Meerbeek, Graduate Student Researcher at Cornell, created soft actuators that allowed them to detect a variety of deformation types, specifically bending and twisting, by inserting 30 optical fibres into silicone foam. Van Meerbeek told Materials World, ‘We wanted to embed a sensing system that could be used to detect a variety of deformation types. We shone a light through the foam and detected the light coming out. Using that optical information in combination with machine-learning, we created a map of light intensity to deformation types, so we could detect different types and magnitude of deformation.’
Two models were used to determine and predict deformation, detailed in the paper, Soft optoelectronic sensory foams with proprioception, published in Science Robotics. Meerbeek said, ‘First, we used a classifier to detect a machine-learning model. That classifier allows us to detect what type of deformation the foam was experiencing – either twisting or bending.’ The foam was distorted over 2,000 times between ‘a few degrees’ to 180˚, and using several machine-learning models including k-nearest neighbours, decision tree, and support vector machine, each used in classification and regression analysis, the Cornell researchers created a system that was able to predict whether the foam had been bent upwards, downwards or twisted, with 100% accuracy. Van Meerbeek continued, ‘Each of the classifier models make different assumptions about your data, but they all worked really well.
‘For regression, in addition to the previous models, we used a multi-layer perceptron called a neural net. There, we saw more variation in accuracy but they all gave what I would consider useable accuracy. K-nearest neighbours worked by far the best, on average within a hundredth of a degree in error – and humans have a few millimetres of error in our fingers, and we can still manipulate objects.’
The Cornell researchers investigated numerous materials, although they have not published all of the information collected. Meerbeek said, ‘In principle, you could use any rubber that allows light to pass through it, so we used a translucent silicone in this case.’ Softer materials, with a lower elastic modulus, may be more sensitive to external pressure and deflect under the same load, and thereafter be more sensitive to detecting the deformations. Similarly, smaller LEDs and photodiodes could be integrated into soft actuators to support an active robotic system.
Bending and twisting deformations were selected for testing due to the ease in establishing laboratory conditions, but the machine-learning capabilities are not limited to these parameters. Meerbeek added, ‘You could also include in that system something that pokes or pinches, as long as you can create the mechanical testing apparatus, but to do that you would need a different way of modelling the deformation.
‘We modelled by looking at bending and twisting only, and you need to look at a 3D model. This is challenging due to the amount of information, and detecting further deformations would need more advanced machine-learning techniques. This would require a more complex mechanical testing setup.’