Martina Zambelli
Martina Zambelli
Embedded Files
I am a research scientist at Google DeepMind. I am interested in robot learning, robot manipulation, multimodal learning.
About
About
I completed my Ph.D. in the Personal Robotics Lab, Imperial College London, in 2017, with Prof. Yiannis Demiris. My Ph.D. thesis focused on multimodal online learning using an iCub humanoid robot. I explored the developmental robotic approach for autonomous discovery of sensorimotor capabilities.
I received my Laurea degree in Automation Engineering from the University of Padova, with a thesis project developed at KTH - Royal Institute of Technology (Stockholm) in the Automatic Control department, on nonholonomic control with performance guarantees, with Prof. Dimos V. Dimarogonas and Yiannis Karayiannidis.
Research
Research
I am interested in dexterous robot manipulation, multimodal and tactile perception, robot learning.
Here's a selection of works (for the full list follow the link at the bottom).
Dexterous manipulation
Coming soon!
Touch representations
Tactile perception is important for manipulation skills. We study how to learn tactile representations from multimodal data in a self-supervised way. The learned representation allows to achieve improved classification across several different objects. [pdf]
Gentle manipulation
Robot manipulation should be safe, both for the robot and its surroundings including objects being manipulated, and to avoid wear and tear. We develop a way to integrate gentleness within reinforcement learning for a dexterous robotic hand. [pdf]
Multimodal representation for prediction and control
We propose a multimodal VAE to learn a sensorimotor representation from own exploration and to achieve control under restricted perception as well as third-person imitation. [pdf]
Online multimodal ensemble for sensorimotor learning
Predicting and acting in continuously evolving environments that intrinsically require the use of different sensory modalities is particularly relevant for autonomous robots. We present a learning architecture based on self-learned multimodal sensorimotor representations. [pdf]
Contacts
Contacts
Email: zambellim [at] google [dot] com
Twitter: @MartinaZambelli
Page updated
Google Sites
Report abuse