Tactile gestures are an integral part of infant development, human communication, and learning. Humans are able to communicate instinctively through tactile gestures bringing down linguistic and cultural barriers, however, robots cannot meet that level of expressiveness yet.
Under the recently finished ROBOSKIN project, I have been working on the development of a novel learning algorithm inspired by the infant cognition model, as introduced by Cohen et al. , which combines the powerful properties of artificial neural networks with the basic elements of RL. Our algorithm is based on a hierarchical structure, whose building blocks are augmented Kohonen Self-Organizing Maps. The bottom level serves as state and action space reduction algorithm, whereas the higher levels encode sequences of winning nodes on the level below producing increasingly abstract representations towards the top of the hierarchy. A complete description of the algorithm should be soon available in the publications section.
Space Telerobotics
Bidirectionally communicating information with robots for both perception and control has been trivial in modern high speed networks. However, time delays, limited bandwidth, and intermittent communication with robots in outer space make real-time control cumbersome. Autonomy has not reached the level to ignore human intervention, therefore the way forward is semi-autonomy, e.g., combination of pre-programmed skills, learned behaviours, and "real-time" control.
In a different configuration, I am interested to use our Hierarchical structure to acquire skill primitives or behaviours. In a real-time control environment, e.g., a human controlling a robot using a Motion Capture suit (like what we did for soccer in the past), we should be able to match already learned skills with recent observations. In the unfortunate event of communication loss, e.g., Mars "turns its back" to Earth or a rover can transmit at most 3 hours per day, our Hierarchy should be able to recover and continue autonomously, until newer directions have been received.
As this work is not part of my PhD, progress is slow; however, I have recently started working on Robonaut's 2 simulator to validate these ideas, on a simple task, such as using a switch on the ISS task board. Click here for high-res.
The Robonaut 2 animation by Georgios Pierris is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License.
Finless Rocket Control
Controlling a finless rocket is a challenging task. In a purely reinforcement learning configuration of our Hierarchical SOM, we expect to learn a stable controller, e.g., for a 4 engine finless rocket. JSBSim is an open source flight dynamics simulator that can provide the environment to learn the controller, however, we still have to develop an interface.