Probabilistic Discriminative Models address the Tactile Perceptual Aliasing Problem

Abstract

In this paper, our aim is to highlight Tactile Perceptual Aliasing as a problem when using deep neural networks and other discriminative models. Perceptual aliasing will arise wherever a physical variable extracted from tactile data is subject to ambiguity between stimuli that are physically distinct. Here we address this problem using a probabilistic discriminative model implemented as a 5-component mixture density network comprised of a deep neural network that predicts the parameters of a Gaussian mixture model. We show that discriminative regression models such as deep neural networks and Gaussian process regression perform poorly on aliased data, only making accurate predictions when the sources of aliasing are removed. In contrast, the mixture density network identifies aliased data with improved prediction accuracy. The uncertain predictions of the model form patterns that are consistent with the various sources of perceptual ambiguity. In our view, perceptual aliasing will become an unavoidable issue for robot touch as the field progresses to training robots that act in uncertain and unstructured environments, such as with deep reinforcement learning.

Pose-and-shear-based tactile servoing

“…Firstly, while the errors are significant across all pose components estimated by both models, they are larger for the shear-related components than the normal contact ones. This could be due to aliasing effects, which are more prevalent during shear motion than normal contact motion; for example, at small contact depths, the tactile sensor is prone to slip under translational shear, which would lead to a similar tactile image for a range of shear values (see Lloyd et al (2021) for an explanation of the effects of tactile aliasing). Secondly, the GDN estimates appear more accurate than the CNN regression estimates, in that the distribution of predicted values is more concentrated around the true values for the GDN model than the CNN regression model, which is consistent with the statistical results presented in Table 1.…”

“…The most obvious limitation of our pose-and-shear estimation methods is the inaccurate CNN regression and GDN model performance on the shear components (Figure 12). We believe that much of this estimation error is due to tactile aliasing (Lloyd et al (2021)), whereby similar tactile images in the training set become associated with very different shear labels. Specifically, when the sensor is sheared sufficiently after contacting a surface, it can slip across the surface to result in similar tactile images for a range of post-contact shear labels.…”

“…This allows us to estimate both the surface contact pose/shear and its associated uncertainty (Figure 3). The motivation for doing this was discussed in our previous work on tactile aliasing (Lloyd et al (2021)): if we know the uncertainty associated with a pose, this information can be used to reduce the error and uncertainty using other system components such as the Bayesian filter we describe in the next section. In our previous work, we considered a Mixture Density Network (MDN) composed of a mixture of Gaussians.…”

“…Two technical challenges had to be overcome to realise a pose-and-shear-based tactile servoing system with a soft high-resolution tactile sensor. Firstly, slippage under shear can cause tactile aliasing (Lloyd et al (2021)) whereby very similar sensor images may become associated with significantly different poses and shears, which can lead to error-prone pose-and-shear estimation that is not suitable for accurate control. To address this, we modified the CNN into a Gaussian-density neural network (GDN) that predicts both the values and the associated uncertainties of the pose-and-shear components, which we then feed through a Bayesian filter to accurately predict pose-and-shear.…”