Abstract:
Tactile perception is important for robotic systems that interact with the world through touch. Touch is an active sense in which tactile measurements depend on the contact properties of an interaction—e.g., velocity, force, acceleration— as well as properties of the sensor and object under test. These dependencies make training tactile perceptual models challenging. Additionally, the effects of limited sensor life and the near-field nature of tactile sensors preclude the practical collection of exhaustive data sets even for fairly simple objects. Active learning provides a mechanism for focusing on only the most informative aspects of an object during data collection. Here we employ an active learning approach that uses a data-driven model’s entropy as an uncertainty measure and explore relative to that entropy conditioned on the sensor state variables. Using a coverage-based ergodic controller, we train perceptual models in near-real time. We demonstrate our approach using a biomimentic sensor, exploring "tactile scenes" composed of shapes, textures, and objects. Each learned representation provides a perceptual sensor model for a particular tactile scene. Models trained on actively collected data outperform their randomly collected counterparts in real-time training tests. Additionally, we find that the resulting network entropy maps can be used to identify high salience portions of a tactile scene.
Discussion:
Fig. 1. A zoomed in view of the leaf token, with the learned model’s entropy heatmap as an overlay. The model has clearly identified the leaf as being an area of particularly high salience on this token.
This publication represents the first year or so of my PhD, and is based on earlier work on sensor-agnostic perception by my co-author Ahalya. The goal for this work was to use active learning — collecting data to maximize learning — to improve the performance of tactile perception models. Tactile is an obvious application for active learning because touch is an inherently near-field sense, with a very diverse range of sensing mechanics. The diverse range of sensing mechanics means that data cannot (easily) be transferred from one sensor to another to create the massive datasets that have powered the ML computer vision revolution. The near-field nature of most tactile sensors, which is to say they sense a small area relative to the size of most objects and spaces, means that it will often not be practical to fully explore an object or space before building a perception model of that environment. That makes the choice of where to explore, and how long, essential to training accurate robust sensor models. The style of active learning we use in the paper, in which the sensor collects data in areas about which the sensor is uncertain, addresses this challenge by “automatically” focusing on areas of high tactile complexity.
My contention, which is not fully fleshed out in this paper, is that tactile complexity is a good proxy for tactile salience. In fact I would go a step further and argue that tactile complexity can often stand in for specific tactile properties during object classification (but that is a story for another paper). A great example of this phenomenon is Figure 1, which shows the uncertainty of a model trained on data from a leaf. Without pre-defining any sense of a leaf’s geometry or properties, the model clearly identifies the organic surfaces as high salience, and provides a pretty good highlight of the leaf’s actual location (plus a bit of margin for the sensor’s diameter).
This aspect of my research was quiescent for a few years while the soft robotics projects took off, but has recently picked up with some exciting work involving new sensors, and new ways of applying the same learning approaches to very conditional spaces and applications.