Hi, I am Manu 👋

My Projects

Visual Navigation using Deep Reinforcement Learning on a Mobile Robot

This work is a collaborative research project withNational Chung Cheng Universityin Taiwan, supervised by ProfessorWen-Nung Lie. For visual navigation of mobile robotics inside the home, the project aimed to test the capabilities of the most advanced Deep Reinforcement Learning technique, namely theAsynchronous Advantage Actor-Critic(A3C). The obtained results were superior to the current scientific literature. Additionally, there were multiple suggestions for extending the neural network architecture using the Self-Attention mechanism, presented in the notorious paperAttention Is All You Need, but the final suggestion was not implemented due to a lack of time and resources.

The code can be found here:GitHub.

Adaptable Impedance learning for UR5 robotic manipulator, as a learning from demonstration system

This work was based on the research paperForce-based variable impedance learning for robot manipulators, and was completed by a team of four. My work was responsible for the GMM and GMR implementation of the system, which took about six months. The project has been tested on an actual UR5, and it functions precisely as described in the paper. Following thislink, actual footage of the project results can be seen, both training and inference results with self adapting impedance reaction.

The code can be found here:GitHub.

LSTM time series prediction AI for RAS systems

In collaboration withBillund Aquaculture, this work proposed, in a team of three, an entirely novel approach for recirculating aquaculture systems time series prediction for distinct key predictors evident from both automatic sensor and manual chemical water tests. The AI was tasked with predicting the total mass of salmon in each reservoir for the following day based on the history of the main predictors. This artificial intelligence is also capable of predicting catastrophic events due to poor decisions made during various actions, such as inadequate nutrition amounts.

Due to NDA some of the code will be missing. But some parts of it can be found here:GitHub.

AI learns to play LUDO

This study proposed a Deep Q-Learning method for creating an AI capable of learning the LUDO game. Multiple variants of the same algorithm were evaluated with varying rewards, opponents, and neural network architectures to determine the mod's optimal algorithm. The final result demonstrates very optimistic results, as well as the algorithm's limitations. The highest results seem to indicate a 61% win rate against three random opponents and a 92% win rate against one random opponent. This algorithm's scope should be an Actor-Critic model, which is anticipated to produce superior results.

The code can be found here:GitHub.

Speaker recognition using Deep Learning

This project contains the code necessary to develop an AI for Speaker Identification using minimal data. The primary objective was to train a rudimentary CNN capable of identifying the speaker in each data sample. The data originate from the podcast "Her Gr Det Godt," and the system was designed to recognize one or both presenters speaking based on 5-second samples. The primary techniques evaluated for data preprocessing originate from this source. The most challenging aspect of the project is its limited database, which posed the greatest challenge when attempting to solve it.

The code can be found here:GitHub.

AI learns to play Flappy Bird

This was one of my first interactions with artificial intelligence and trainable programs. In this work, I attempted to replicate aYouTube-sourcedconcept for an automaton, and it served as a precursor to what is currently my greatest passion. This bot's primary learning algorithm, NEAT (Neuroevolution of Augmenting Topologies), is fundamentally distinct from gradient methods, which I continued to use later, but it taught me many valuable lessons and inspired me to pursue this field even further.

The code can be found here:GitHub.

Future Talent Denmark

This program was created to facilitate connections between Danish firms and international labor force by assisting Danish firms in expanding their talent pipeline by bridging the divide between firms and international talent in the country's academic institutions. It lasted six months and definitely helped me adjust to the new job market.

The certification can be found here:Link.

1st place SDU Hackathon 2021 sponsored by Danfoss

This was one of my most outstanding accomplishments, as our team of five at the University of Southern Denmark in Sonderborg won first place in a highly competent and competitive environment. Our team designed and presented a prototype for a free-range farm egg harvester that persuaded the jurors and the general public that it would be the best idea for a sustainable solution that would reduce highly demanded hard labor. The following links provide additional information:link1,link2, andlink3. As a consequence, we were awarded a prize of 15,000 Danish Krone from Danfoss.