The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Relational, Object-Centric Agents for Completing Simulated Household Tasks with Wilka Carvalho - #402
Aug 20, 2020
In a fascinating discussion, Wilka Carvalho, a PhD student at the University of Michigan, delves into his research on AI agents designed for completing household tasks. He explores the challenges of object interaction and how his Relational Object Model Learning Agent (ROMA) tackles these issues. Carvalho emphasizes the importance of representation learning and real-world data in creating effective agents. He also shares insights on achieving high sample efficiency and the methodologies for training robotic agents in complex environments.
AI Snips
Chapters
Transcript
Episode notes
Path to AI
- Wilka Carvalho's interest in studying the brain led him to physics, then to machine learning through reinforcement learning.
- He believes algorithms and models are key to understanding intelligence, pursuing AGI research in his computer science PhD program.
Algorithms of Intelligence
- Wilka sees machine learning as a path towards understanding intelligence, even artificial general intelligence (AGI).
- He believes algorithms and models are the fundamental basis of intelligence and behavior, not just biophysical mechanisms.
Generalizable AI for Human Environments
- Wilka Carvalho focuses on building AI agents that learn generalizable knowledge and skills in human-like environments, like homes or kitchens.
- This research involves representation learning and reinforcement learning, advised by Hong Lok Lee and Satinder Singh.