Joe Saunders

In the above titled paper (ibid., vol. 1, no. 3, pp. 187-195, Oct. 09), the acknowledgement to financial support was incompletely displayed. The correct acknowledgement is presented here.

Research into robotic social learning, especially that concerned with imitation, often focuses at differing ends of a spectrum from observational learning at one end to following or matched-dependent behaviour at the other. We study the implications and differences that arise when carrying out experiments both at the extremes and within this spectrum. Physical Khepera robots with minimal sensory capabilities are used, and after training, experiments are carried out where an imitating robot perceives the dynamic movement behaviours of another model robot carrying a light source. It learns the movement behaviour of the model by either statically observing the model, dynamically observing the model or by following the model. It finally re-enacts the learnt behaviour. We compare the results of these re-enactments and illustrate the differences and trade-offs that arise between static observational and reactive following learning methods. We also consider circumstances where, for this robotic embodiment, dynamic observation has both advantages and disadvantages when compared to static observation. We conclude by discussing the implications that arise from using and combining these types of social learning.

Imitative learning and learning by observation are social mechanisms that allow a robot to acquire knowledge from a human or another robot. However to be able to obtain skills in this way the robot faces many complex issues, one of which is that of finding solutions to the correspondence problem. Evolutionary predecessors to observational imitation may have been self-imitation where an agent avoids the complexities of the correspondence problem by learning and replicating actions it has experienced through the manipulation of its body. We investigate how a robotic control and teaching system using self-imitation can be constructed with reference to psychological models of motor control and ideas from social scaffolding seen in animals. Within these scaffolded environments sets of competencies can be built by constructing hierarchical state/action memory maps of the robot's interaction within that environment. The scaffolding process provides a mechanism to enable learning to be scaled up. The resulting system allows a human trainer to teach a robot new skills and modify skills that the robot may possess. Additionally the system allows the robot to notify the trainer when it is being taught skills it already has in its repertoire and to direct and focus its attention and sensor resources to relevant parts of the skill being executed. We argue that these mechanisms may be a first step towards the transformation from self-imitation to observational imitation. The system is validated on a physical pioneer robot that is taught using self-imitation to track, follow and point to a patterned object.

Teaching a robot new skills may require that the teacher scaffolds the teaching experience appropriately. However, due to inherent assumptions made by a human teacher the scaffolding process may in some circumstances fail to effectively teach the robot. Here we illustrate this issue in two simple robot teaching exploratory studies and examine the assumptions made by the teacher when teaching the robot. In the first study the human teacher had to reason about robot perceived states in order to provide suitable teaching. In the second study the human teachers had to understand the perceptual constraints of the robot based on the instructions given beforehand by the experimenter and subsequently adapt the guidance given. The results suggest that although the two tasks are quite distinct in their level of complexity a common thread can be observed: people tend to underspecify their teaching. It seems that steps of the explanation are assumed to be known and skipped or not even considered at all. We reflect on the possibility that one of the major challenges in designing robots that are capable interaction partners in these teaching situations is to be able to make them communicate their internal state and current capabilities effectively. Furthermore, we also reflect on designing appropriate behavioral primitives for the robot, corresponding implications on the level of task description and for benefiting from human teaching.

One way in which a robotic system might be considered 'intelligent' may be an ability to learn and then generalise its learnt abilities in situations not previously encountered. Thus the ability to learn may be a prerequisite in allowing a robot to be adaptable when coping with changing environments, be useful when dealing with changing user requirements and expectations as well as being in itself a mechanism which supports the idea of intelligence and intelligent behaviour. One of the aims of the robotics research presented here is to study adaptive mechanisms which support learning with a focus on supervised learning via the interaction between a human teacher and a robot learner [6]. This interaction is based upon learning via imitation from observation of others as well as self-imitation, where the learner learns by reproducing actions it has perceived from another manipulating its body in order to highlight affordances and effectivities available for action [7]. In this talk I will suggest that the psychological notion of 'ideomotor theory' may be a good starting point from which to construct adaptable robots which support learning from imitation. In doing this I will explain the ideas behind ideomotor theory, ranging from the initial work of Lotze [3] and James [2] to more recent research by Prinz [5] where the theory is extended to support imitation. I will contrast and compare ideomotor theory with other prevailing theories of imitation, such as Active intermodal matching [4] and Associative Sequence Learning [1] and suggest that the ideomotor approach may have more to offer the roboticist in conceiving learning systems. Finally I will demonstrate how these ideas can be put into practise with some examples of scaffolded teaching used to teach physical robots an object following task.

An evolutionary predecessor to observational imitation may have been self-imitation. Self-imitation is where an agent is able to learn and replicate actions it has experienced through the manipulation of its body by another. This form of imitative learning has the advantage of avoiding some of the complexities encountered in observational learning such as the correspondence problem. We investigate how a system using self-imitation can be constructed with reference to psychological models of motor control including ideomotor theory and ideas from social scaffolding seen in animals to allow us to construct a robotic control system. The system allows a human trainer to teach a robot new skills and modify existing skills. Additionally the system allows the robot to notify the trainer when it is being taught skills it already possesses. We argue that this mechanism may be the first step towards the transformation from self-imitation to observational imitation. We demonstrate the system on a physical Pioneer robot with a 5-DOF arm and pan/tilt camera which is taught using self-imitation to track and point to coloured objects