Joose Rajamäki

Joose Rajamäki · Mon, 02.10.2017

I attended the Symposium of Computer Animation (SCA) and SIGGRAPH, which were co-located in Los Angeles. In SCA I presented the paper “Augmenting Sampling Based Controllers with Machine Learning”. I am the first author of the paper and it will be included to my PhD. In SIGGRAPH I had co-authored the paper “Discovering and Synthesizing Humanoid Climbing Movements”.

The paper “Augmenting Sampling Based Controllers with Machine Learning” extends the existing sampling based controller presented in our research group’s previous paper “Online Control of Simulated Humanoids Using Particle Belief Propagation”. The probabilistic framework of the paper is simplified by replacing it with a Monte Carlo Tree Search type algorithm. Furthermore, the sampling process is augmented – as the title suggests – with machine learning. A central observation of the paper is that the combination of a quick nearest neighbor learner and a more gradually learning neural network yields better results than either learner alone.

The paper “Discovering and Synthesizing Humanoid Climbing Movements” separates the problem of climbing movements to two hierarchies. First, there is a graph search between the different hold configurations that the climber may take. Secondly, there is a low-level sampling based controller that tries to implement the transitions selected by the graph search. This combination allows long term planning with highly dynamic movements.

The conferences themselves were quite different and it was nice to see a small conference (SCA) where you got to chat with all the participants and a big conference (SIGGRAPH) where you’d see a wide variety of research, art and technology. Also, Los Angeles in itself was quite an experience.