Reproduce “Probabilistic Discriminative Models Address the Tactile Perceptual Aliasing Problem” on OpenResearch

About this paper

John Lloyd, Yijiong Lin, Nathan F. Lepora

This study introduces the concept of Tactile Perceptual Aliasing, where distinct physical interactions yield similar sensor outputs, and addresses it using probabilistic discriminative models. Mixture Density Networks (MDNs) are demonstrated to provide improved predictive accuracy and robust uncertainty quantification, enabling the identification and characterization of ambiguous tactile data in robotic sensing.

Key insights

• Perceptual aliasing significantly degrades the predictive accuracy of conventional discriminative regression models in tactile sensing, resulting in higher Root Mean Square errors and wider prediction dispersions.
• Probabilistic discriminative models, like MDNs, can effectively model the multi-modal nature of aliased tactile data, providing not only predictions but also interpretable uncertainty measures that directly reflect data ambiguity.
• The uncertainty estimates derived from MDNs provide a clear, visual correspondence with known sources of perceptual aliasing, such as light contacts or flat surfaces, offering deeper insights into the tactile perception process.

Summary by alphaXiv.