Author Image

Hi, I am Simone Zamboni

I am a Resear|

Simone Zamboni

Machine Learning Research Engineer at Embark Studios

Currently Machine Learning Research Engineer at Embark Studios developing language models for video game characters. Previously Machine Learning Developer and Master’s Degree in Autonomous Systems at KTH. Accomplishments include publishing a journal research article based on my master’s thesis at SCANIA and presenting advancements in language models for video game characters at conferences.

Skills

Experiences

1
Embark Studios

March 2022 - Present

Stockholm

Video game company founded in 2018 by industry veterans from DICE

Machine Learning Research Engineer

March 2022 - Present

Responsibilities:
  • Researching how to use large language models to make video game characters come alive
  • Keeping up to date with the latest research
  • Fine-tuning language models with python and PyTorch
  • Help with the deployment of language models by preparing Docker images
  • Develop on the game client which is written in Rust.
Substorm

November 2020 - February 2022

Stockholm

Substorm AB is consultancy company specialized in AI

Machine Learning Developer

November 2020 - February 2022

Responsibilities:
  • Developed and implemented Natural Language Processing (NLP) solutions using Python, PyTorch, TensorFlow/Keras, Scikit-learn, Numpy, and Pandas.
  • Managed the deployment processes with AWS, Docker, Azure Pipelines, and Terraform.
  • Supervised a master’s thesis student from KTH on a deep Reinforcement Learning project.
2
3
Scania.

Jun 2020 - June 2020

Södertälje

Machine Learning Engineer Intern

Jun 2020 - June 2020

Responsibilities:
  • Conducted a master’s thesis project at SCANIA Autonomous Driving Research, focusing on pedestrian trajectory prediction using deep neural networks.
  • Achieved state-of-the-art performance on a publicly available dataset through innovative research methods.
  • Published thesis findings as an article in the journal Pattern Recognition (available at https://doi.org/10.1016/j.patcog.2021.108252).

Education

KTH
2018-2020
Double Master Degree in Autonomous Systems
Description:
Double Master Degree between the University of Trento, for the first year, and the Royal Institute of Technology (KTH), for the second year, focused on Artificial Intelligence, Robotics and Entrepreneurship. Topics addressed include: machine learning, deep learning, reinforcement learning, robotic design, robotic path planning and start-up business development. This program, supported by EIT Digital, a European institution that aims to create highly technical figures (in the robotics and AI fields) with also business knowledge applicable to the startup world. Highest possible exit grade (110 with honors out of 110).
University of Trento, Italy
2018-2020
Bachelor's Degree in Information and Business Organization Engineering
Description:
Bachelor Degree focused on Computer Science, with topics like Object Oriented Programming (in C++), UML diagrams, database design and query with SQL, networking, Javascript development and Python development. A part of my study plan was also focused on business organization: I studied agile techniques for software development, process management and business strategy. Highest possible exit grade (110 with honors out of 110)

Publications

Pedestrian trajectory prediction with convolutional neural networks
Pattern Recognition Jan 2022
Simone Zamboni Zekarias Tilahun Kefato Sarunas Girdzijauskas Christoffer Norén LauraDal Col

Predicting the future trajectories of pedestrians is a challenging problem that has a range of application, from crowd surveillance to autonomous driving. In literature, methods to approach pedestrian trajectory prediction have evolved, transitioning from physics-based models to data-driven models based on recurrent neural networks. In this work, we propose a new approach to pedestrian trajectory prediction, with the introduction of a novel 2D convolutional model. This new model outperforms recurrent models, and it achieves state-of-the-art results on the ETH and TrajNet datasets. We also present an effective system to represent pedestrian positions and powerful data augmentation techniques, such as the addition of Gaussian noise and the use of random rotations, which can be applied to any model. As an additional exploratory analysis, we present experimental results on the inclusion of occupancy methods to model social information, which empirically show that these methods are ineffective in capturing social interaction.

Machine Learning Autonomous Driving Deep Learning
Details

Recent Posts

Hero Image
Slice of AI 3 - Bridging Worlds - Recent advances in Multi-modal AI

Paper presentation and discussion session, organized by Stockholm AI and Silo AI, with the topic of Multi-modal AI. This session was moderated by Pier Luigi Dovesi and I was the main speaker. I started the presentation with an introduction of multi-modal models using the transformer architecture to then deep dive into two recent papers: CogVLM: Visual Expert for Pretrained Language Models (https://lnkd.in/dPaB946G) mPLUG-Owl2: Revolutionizing Multi-modal Large Language Model with Modality Collaboration (https://lnkd.

Thursday, April 11, 2024 | 1 minute Read
Hero Image
Everything Procedural 2024 Talk

Presented at the Everything Procedural 2024 conference in Breda(NL) how we are using AI at Embark Studios to make video game characters come alive. Event URL: https://www.everythingprocedural.com/ Presentation recording: https://www.youtube.com/watch?v=aOkw4aCo9Ro We would like the interactions with our characters to be unique and flexible, so we are using Language Models to do that. We create a text representation of the world around our characters and we use it together with the user input to prompt a fine-tuned language model to generate a response.

Friday, June 16, 2023 | 1 minute Read
Hero Image
Dreamhack 2023 Talk

The presentation was about how we are using AI at Embark Studios to make video game characters come alive. We are using Reinforcement Learning to make our characters move in a physically-based environment, so that we don’t have to make complicated animations like it is traditionally done in video games. We also want interactions with our characters to be unique and flexible, so we are using Language Models to do that.

Friday, June 16, 2023 | 1 minute Read

Projects

Detecting pedestrians at night
Detecting pedestrians at night
Developer 2019

Implementation of a deep learning detection model to detect pedestrians in images taken at night from the NightOwls dataset. The model was implemented with Python and the Keras API for TensorFlow and trained on the Google Cloud Platform. This project was for the course of Scientific Reading and Writing at KTH.

university deep-learning
Generate faces with Generative Adversarial Networks
Generate faces with Generative Adversarial Networks
Developer 2019

Implementation of a Generative Adversarial Network (GAN) that generates faces with specified features. The network was trained on the CelebA dataset and was implemented using Pytorch. It was the project for the course of deep learning at the University of Trento.

university deep-learning
Path planning for wheeled robot
Path planning for wheeled robot
Developer 2019

Plan a path for a wheeled robot: given an image taken from the top of a robot in an environment with obstacles and a goal, the objective of the project was to plan a path to reach the goal, avoiding the obstacles. The project was written in C++ with the image processing part done with OpenCV. Project for the course of Laboratory of Applied Robotics at the University of Trento.

university
NIPS2019 paper reproduction
NIPS2019 paper reproduction
Developer 2019

Project for the advanced deep learning course at KTH where the task was to choose a paper from NIPS 2019 and reproduce it. The chosen paper was about transfer learning: the objective was to train a network from another already trained network (the “teacher”) without any of the data the teacher network was trained with. The name of the paper is: “Zero-shot Knowledge Transfer via Adversarial Belief Matching”, it was re-implemented using Python and Pytorch and trained on the Google Cloud Platform.

university deep-learning
Detecting pedestrians at night
Detecting pedestrians at night
Developer 2019

Implementation of a deep learning detection model to detect pedestrians in images taken at night from the NightOwls dataset. The model was implemented with Python and the Keras API for TensorFlow and trained on the Google Cloud Platform. This project was for the course of Scientific Reading and Writing at KTH.

university deep-learning