Introspective Methods for Reliable Autonomy

Abstract - TBA

Following characterizations of introspection in philosophy, there cannot be introspection in robots. (And probably not in humans either – we probably just have the sensation of introspection.) I will talk about philosophy briefly, hopefully helping us to articulate clearly what we might mean by introspection in robotics. Towards this goal, the workshop seeks answers the question of “How to assess the quality of internal models, methods and sensor data, used by robots and how to alter their behavior upon this information?” I am going to argue a) that assessing quality of internal models and sensor data is ubiquitous in robotics (think: Kalman filter) and b) that altering behavior based on this information is equally ubiquitous (think: control based on Kalman filter). But that is not really what we mean, right? So what then do we actually mean by introspection? I will propose that this has something to do with hierarchical processing of sensor information in perception for action (or interactive perception). I will also propose a way of designing “introspective” systems and give concrete examples.

Recently, Autonomous Learning methods have become very popular in robotics, because they tend to learn more efficiently and more adaptively by integrating decision making into the learning process. For robot perception, this means that data samples that are hard to classify are selected actively to be labeled by the supervisor for re-training. This is also known as Active Learning. In this presentation, I will show example applications of the Active Learning framework to 3D object classification, and I will argue that for this it is necessary that the classifier be introspective. More in detail, I will motivate the notions of under- and overconfidence and how this relates to introspectiveness. Furthermore, I will show that in particular ensemble learning methods such as bagging and boosting are very useful in this context, because either they already tend to be introspective or they can be modified such that they are. As a result, Active Learning produces less label queries and provides a better classification accuracy. Based on experiments on benchmark data sets I will show how this can be used to learn 3D objects more efficiently and more accurately, even if the data is given in streams.

Abstract - TBA

The development of autonomous systems seems to be at a critical point.

For applications where these systems have to operate in more complex environments reaching more complex goals, a number of problems are encountered.

The systems need a lot of computational power and memory, they have difficulty explaining themselves, they can behave in an unpredictable or unreliable way, they do not use their resources effectively, they are not robust to failures and they have difficulty to understand and act according to the law.

Because of these problems also test and evaluation of these systems is problematic.

It is understood that introspective methods can help and I will argued that these introspective methods can be regarded as a form of self-awareness of the system; the capability of doing self-assessment and self-management.

By presenting various system models it is discussed how self-awareness could be implemented and also on the basis of some examples where ‘self’ may refer to.

Finally I will touch upon the problems regarding ethical and lawful behavior.

Artificially Intelligent (AI) robots are increasingly finding many of their critical capabilities dependent on complex Machine Learned (ML) models. Unlike planners, hand coded controllers or mathematically derived models, there is considerable variation in the explainability of different ML techniques, from completely opaque, black-box models to transparent, rule based models. When selecting a ML technique, in addition to on-task performance, there needs to be careful consideration of the explanatory capabilities of the ML techniques and the explanatory demands of the application, both for the agents' own introspection and for presentation to humans.

Explainable Artificial Intelligence (XAI) is becoming topical as a study of the explainability of various ML techniques in particular, and AI techniques more generally. In this talk we will discuss some of the implications of these variations in explainability and propose a categorisation of explainability from the perspective of matching up the robotic need for introspection and human-robot interaction (HRI) with the explanatory capabilities of various ML techniques.

Abstract - TBA