Active exploration in the high-dimensional data of an artificial skin
Final Report Abstract
This project explored handling and interpretation of data from a novel tactile sensor provided by DLR. Contrary to existing tactile sensors, it has the advantage of being flexible and can thus be used to coat robotic grippers and fingers. Consequently this tactile sensor is the robotic equivalent of a biological skin. The working principle of the skin is simple. It consists of two layers of a conductive silicone compound separated by weakly conductive polymer compound. When pressure is applied the polymer layer gets deformed which results in a change of the skin's electrical properties, namely resistivity and capacity. The flexibility comes at a cost, however. The employed polymer behaves highly non-linear with respect to the applied force and its behavior changes depending on previously applied forces. It can be said that the skin has a memory effect, also referred to as hysteresis. Hence the sensed electrical properties are challenging to correlate to the applied force. Furthermore due to variances in the production process each sensing element ("taxel") of the skin behaves differently, making the development of an explicit model problematic. At project start it was expected, that the skin would be available from DLR with measurement electronics that provide a signal which varies continuously with the applied pressure. In reality, though, the provided readout electronics were only able to discriminate if a force above a certain threshold was present or not. Therefore a new version of the measurement electronics based on alternating current (AC) stimulation and lock-in amplification had to be developed within this project. To our knowledge this resulted in the first high-precision readout system for a flexible artificial skin that is resilient to electromagnetic interference commonly present in a robotic environment. Still to make use of the skin as a pressure sensor, the measured change in electrical capacity has to be converted into force. For this purpose we developed an algorithm that can learn to do so by applying test forces to a taxel of the skin and observing how the electrical properties change accordingly. The algorithm thus performs calibration of the skin without the necessity for a physical model. Because it also learns to model the internal state ("memory") of the polymer and its evolution over time, the algorithm automatically compensates for hysteresis. It has been shown that in a biological brain information is combined additively as well as multiplicatively. In contrast to that most artificial neural networks in today's research and commercial application use additive interactions only. While this approach has proven successful in practice and powers today state-of-the-art image and voice recognition systems, in theory a combination of additive and multiplicative interactions is substantially more powerful and can solve more problem classes compared to the additive-only counterpart. One challenge to employing these hybrid additive-multiplicative-networks is that there was no efficient method of determining what operation (addition or multiplication) a particular neuron should perform. The reason for this difficulty is that addition and multiplication are two fundamentally different operations with no known continuous transition between them. During this project we proposed a mathematical framework that defines these continuous transitions and thus allows neurons to learn the operation they should perform during training of the artificial neural network using currently established methods without considerable additional overhead.
Publications
- "Computing grip force and torque from finger nail images using Gaussian processes." In Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on, pp. 4034-4039. IEEE (2013)
Urban, Sebastian, Justin Bayer, Christian Osendorfer, Goran Westling, Benoni B. Edin, and Patrick van der Smagt
- "Training neural networks with implicit variance." In Neural Information Processing, pp. 132-139. Springer Berlin Heidelberg (2013)
Bayer, Justin, Christian Osendorfer, Sebastian Urban, and Patrick van der Smagt
- On fast dropout and its applicability to recurrent networks
Bayer, Justin, Christian Osendorfer, Daniela Korhammer, Nutan Chen, Sebastian Urban, and Patrick van der Smagt
- "Estimating finger grip force from an image of the hand using Convolutional Neural Networks and Gaussian Processes." In Robotics and Automation (ICRA), 2014 IEEE International Conference on, pp. 3137-3142. IEEE (2014)
Chen, Nutan, Sebastian Urban, Christian Osendorfer, Justin Bayer, and Patrick van der Smagt
(See online at https://doi.org/10.1109/ICRA.2014.6907310) - "Sensor calibration and hysteresis compensation with heteroscedastic Gaussian processes." IEEE Sensors Journal (Volume: 15 , Issue: 11 , Nov. 2015) pp 6498-6506
Urban, Sebastian, Marvin Ludersdorfer and Patrick van der Smagt
(See online at https://doi.org/10.1109/JSEN.2015.2455814) - A Neural Transfer Function for a Smooth and Differentiable Transition Between Additive and Multiplicative Interactions
Urban, Sebastian, and Patrick van der Smagt
