Learning efficient optimal periodic behaviours for mechanical systems #neuralODE
30 Decemeber 2022
Have you ever thought about the problem of learning efficient periodic behaviours for robots (e.g. pick-and-place or flapping)? Well, we have!
In our most recent work, we used a neural network (in particular a neural ODE) to shape the (close-loop) behaviour of our system such that its natural evolution corresponds to the desired periodic movement. [article]
We will soon release our code too!
The state representation learning zoo #unsupervisedlearning
25 November 2022
A few months ago, I embarked on a quest with the goal of writing a complete and understandable review on the field of "State Representation Learning" for experts and non-experts in the field. State Representation Learning is the problem of recovering low-dimensional and meaningful state representations from high-dimensional data. Recovering the state vector is crucial for unveiling dynamics and control of robots and dynamical systems.
It was great learning experience (it was my first review paper) and I am really happy with the results! Enjoy the reading! Get ready because it is a long paper ;) [article]
I have also developed and released the code comparing 18 of the different methods on the problem of learning a compact state representation for a pendulum from high-dimensional observations (i.e. RGB images) with and without visual distractors. [code]
3x IRONMAN 70.3 finisher and new PB
25 Septemer 2022
What a day! For the 3rd time, I am a finisher of the IRONMAN 70.3 Italy and I scored a new personal best (pb) too!
It only ;) took me 4:49:03h to complete 1.9km of swimming in the sea, 90km of cycling, and 21.1km running for a total of ~113km (70.3miles)!
Dimensionality reduction with autoencoders and kernel-based methods #deepkernellearning
1 September 2022
So happy to have completed my first work as a postdoc!
In our work [here], we study the problem on dimensionality reduction, model-order reduction, and uncertainty quantification for dynamical systems from high-dimensional and noisy measurements! All these problems can be tackled jointly with a smart combination of deep neural networks and gaussian processes, i.e. the deep kernel learning.
We have released our code too! [code]
First race of the triathlon season!
15 June 2022
The first race of the season is always the toughest, no matter distance and preparation! However, it was a great fun to push hard again after the winter break :)
Back to physical conferences #ICTOPEN
10 April 2022
After two years, I finally got the chance to physically present my work on "State and Action Representation learning for Reinforcement Learning" [article][presentation] to a physical audience (and not to my laptop screen)! I was a bit stress to do that after so long, but only after a few seconds, the excitement too over!