Miguel Vasco

I am a Postdoctoral fellow at the Division of Robotics, Perception and Learning, at KTH Royal Institute of Technology, where I work on multimodal representation learning and reinforcement learning. My postdoctoral advisor is Danica Kragic.

I am a former intern at Sony AI and a RSS Pioneer. I am also the co-creator of the Talking Robotics podcast.

I have a PhD in Computer Science from Tecnico, University of Lisbon, where I was supervised by Ana Paiva and Francisco S. Melo. I have an MSc and BSc in Engineering Physics from Tecnico, University of Lisbon.

CV  /  Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

I'm interested in multimodal representation learning and reinforcement learning. Here you can find some selective publications. For a complete list check out this link.

Centralized Training with Hybrid Execution in Multi-Agent Reinforcement Learning

We introduce hybrid-execution, a novel paradigm for cooperative multi-agent systems where agents can passively share their observations at execution time, conditioned on an imperfect communication system.

Geometric Multimodal Contrastive Representation Learning

We show how contrastive learning can be used to learn multimodal representations that are robust to missing information at test time. We evaluate our approach in supervised, unsupervised and reinforcement learning tasks.

Perceive, Represent, Generate: Translating Multimodal Information to Robotic Motion Trajectories

We design a robotic system that is able to perceive multimodal instructions from human-users to execute handwriting tasks.

How to Sense the World: Leveraging Hierarchy in Multimodal Perception for Robust Reinforcement Learning Agents

We introduce MUSE, a multimodal representation learning module for RL agents that allows the robust execution of tasks, regardless of the available modalities at execution.

Leveraging Hierarchy in Multimodal Generative Models for Effective Cross-modality Inference

We investigate the problem of cross-modality inference (CMI) in multimodal generative models and show the potential of considering hierarchical representation spaces. We contribute with a novel multimodal MNIST-like dataset.

Playing Games in the Dark: An Approach for Cross-modality Transfer in Reinforcement Learning

We propose a framework that allows RL agents to transfer policies across different modalities (even to unseen ones during policy training!). We also contribute multimodal Atari environments.

Learning Multimodal Representations for Sample-efficient Recognition of Human Actions

We propose a representation for sample-efficient motion recognition of human users in a household environment.

Teaching

FDD3359 Reinforcement Learning, PhD, KTH Royal Institute of Technology - Spring 2024

DD2430 Project Course in Data Science, MSc, KTH Royal Institute of Technology - Fall 2023

Planning, Learning and Intelligent Decision-Making, MSc, Instituto Superior Técnico, University of Lisbon - Fall 2022, Fall 2021

Computation and Society (AI Ethics), BSc, Instituto Superior Técnico, University of Lisbon - Spring 2019, Spring 2020

Students

Current PhD Students

Nona Rajabi, KTH Royal Institute of Technology (co-supervised with Danica Kragic and Mårten Björkman)

Farzaneh Taleb, KTH Royal Institute of Technology (co-supervised with Danica Kragic and Mårten Björkman)

Bernardo Esteves, Instituto Superior Técnico, University of Lisbon (co-supervised with Francisco S. Melo)

Past Master Students

Afonso Fernandes, Instituto Superior Técnico, University of Lisbon (co-supervised with Francisco S. Melo)

Bernardo Esteves, Instituto Superior Técnico, University of Lisbon (co-supervised with Francisco S. Melo)

Fábio Vital, Instituto Superior Técnico, University of Lisbon (co-supervised with Alberto Sardinha and Francisco S. Melo)

Design and source code from Jon Barron's website