IPAB Workshop - 14/03/2019

Speaker: Ian Mason

Talk title: Few-shot Learning of Homogeneous Human Locomotion Styles

Talk abstract:  I will talk about  recent work in which we experimented with training neural networks for character animation to model new styles of locomotion (e.g. old, duck footed) when given only a short reference animation of the style. We propose a transfer learning approach making use of the canonical polyadic tensor decomposition to reduce the amount of parameters required for learning each new style. Given a pretrained character controller in the form of a Phase-Functioned Neural Network for locomotion, our system can quickly adapt the locomotion to novel styles using only a short motion clip as an example, which both reduces the memory burden at runtime and facilitates learning from smaller quantities of data.

Speaker: Calum Imrie

Talk Abstract: Swarm robotics usually defines a collection of individual agents that are limited in sensory and actuator capabilities, and requires the collective to succeed in a task. This means that they rely strongly on local information and interaction with other robots, it is crucial to set up the individual sensing and communication capabilities appropriately. However, in massive swarms often inexpensive robots are used which have limited bandwidth and information processing capabilities, so that it appears to be reasonable to compensate the limitations by relatively large swarms and versatile, robust control strategies which are often found in biological examples. This talk will present the current study which aims to the goal of enabling swarms of simple robots to form as a collective complex patterns.  Specifically we look at spatio-temporal patterns, as they are of potential use in applications beyond classical robotics, in particular in nanorobotics, but also in agriculture, space and marine applications.

Many swarm approaches are based on self-organisation methods, which are usually nature inspired. Our research is based on spatially self-organised natural phenomena that can be modelled by Reaction Diffusion systems, which are known to produce the so-called Turing Patterns. We show here that these patterns are not restricted to physical and biological systems, but can also be used to shape the global structure of a robot swarm in a controllable manner.

Speaker: Marija Jegorova

Talk title: Generating diverse exploration policies with GANs

Talk Abstract: Classic reinforcement learning algorithms, like DDPG, generally suffer from insufficient exploration when sparse or deceptive rewards are involved. They also usually focus on finding a single optimal solution which might prove inefficient for the new environments with changing topology and targets. We propose an adversarial policy generation method for efficiently exploring the new environments based on examples of successful exploration in training domains.