Research Overview

My research interests lie at the intersection of Machine Learning, Robotics, and the pursuit of Physical Artificial Intelligence (AI). I am mostly interested in Reinforcement Learning (RL) and its applications to robot learning systems that can reason, adapt, and act robustly in the real world.

My long-term goal is to contribute to the development of generalist robots capable of robust, adaptive, and lifelong learning behavior across diverse real-world tasks — bridging the gap between learning in simulation and reasoning in the physical world.


Continual RL ➤

My current research direction focuses on improving generalization, sample efficiency, and lifelong adaptability in RL. I am particularly interested in developing algorithms that can learn continuously from experience, adapt to new contexts and users, and leverage human feedback — without explicit task segmentation or manual reward design.

I aim to study methods that go beyond mitigating catastrophic forgetting, focusing instead on forward transfer — enabling past knowledge to accelerate future learning — and the loss of plasticity. My long-term aim is to build task-agnostic agents that evolve continuously in dynamic environments without the need for resetting or explicit supervision.

Causal Reasoning in RL ➤

I am very interested in exploring how causal representation learning can structure the latent spaces of RL agents, improving generalization, explainability, and data efficiency. This involves learning disentangled, causally grounded state representations that allow for transfer across tasks and environments — moving closer to generalist robotic intelligence.

Equivariance and Symmetry in RL ➤

I am also strongly attracted to equivariant neural architectures and how they can encode the underlying symmetries of physical environments (e.g., SE(3) transformations in manipulation). Such inductive biases can dramatically improve generalization across viewpoints and configurations, and are especially promising when combined with continual or causal RL frameworks.

Human-in-the-loop RL ➤

I am increasingly interested in Preference-based RL (PbRL) and Inverse RL (IRL) paradigms, where robots learn reward functions from human corrections, demonstrations, or natural language feedback. This human-in-the-loop perspective is crucial for developing adaptive, personalized, and trustworthy robot behaviors.

Involving Language and Touch in Robot Learning ➤

I am interested in how RL agents can benefit from rich multimodal perception, integrating language, vision, proprioceptive and tactile feedback to improve learning and interaction. Language provides a high-level, interpretable signal that can guide exploration and shape task understanding, while tactile sensing allows agents to perceive and respond to forces, contact, and compliance in real-world environments.

This multimodal perspective enables agents to acquire skills more robustly, reason about complex interactions, and align with human intentions in dynamic, unstructured environments.

Ongoing projects