My name is Robby Costales and I am a machine learning researcher focused on developing adaptive, open-ended agents that can explore and learn across complex, evolving task distributions. I am currently completing my PhD in computer science at the University of Southern California, advised by Prof. Stefanos Nikolaidis in the ICAROS lab.


Experience

Google DeepMind (Gemini)

Google DeepMind (Gemini)

💼 Research Scientist Intern (Summer 2025)

University of Southern California (Viterbi School of Engineering)

University of Southern California (Viterbi School of Engineering)

🎓 Ph.D. Candidate in Computer Science (present)

Google Research (Brain Team)

Google Research (Brain Team)

💼 Student Researcher (2022)

Columbia University (Fu Foundation School of Engineering)

Columbia University (Fu Foundation School of Engineering)

🎓 B.S. in Computer Science - Intelligent Systems (2020)

Bard College at Simon's Rock

Bard College at Simon's Rock

🎓 A.A. (2017) and B.A. in Computer Science (2020)


Selected publications

Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity

Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity [Site] [Paper] [Code]

R Costales, S Nikolaidis

Neural Information Processing Systems (NeurIPS) 2024

An evolutionary approach for generating diverse tasks to train adaptive agents in open-ended simulators.

ALMA: Hierarchical Learning for Composite Multi-Agent Tasks

ALMA: Hierarchical Learning for Composite Multi-Agent Tasks [Paper] [Code]

S Iqbal, R Costales, F Sha

Neural Information Processing Systems (NeurIPS) 2022

A general learning method for leveraging structured multi-agent tasks, resulting in sophisticated coordination behavior and outperforming competitive MARL baselines.

Possibility Before Utility: Learning And Using Hierarchical Affordances

Possibility Before Utility: Learning And Using Hierarchical Affordances [Paper] [Code]

R Costales, S Iqbal, F Sha

🏅 (Spotlight) Int. Conf. on Learning Representations (ICLR) 2022

A hierarchical reinforcement learning (HRL) approach that learns a model of affordances to proune impossible subtasks for more effective learning.