Robotics @ MLCS

Research

We work on robot learning — developing algorithms that enable robots to learn and act in the real world. Our work spans the following directions.

Data-efficient Robot Learning

We exploit the geometric structure of robotics tasks — symmetry and equivariance (e.g. SE(3)) — so that models learn far more data-efficiently and generalize from only a handful of demonstrations.

SE(3)-Equivariance Data Efficiency Geometric Deep Learning

Robot Foundation Models

To generalize across diverse real-world tasks, we leverage foundation models pre-trained on large-scale data from many domains — vision, language, and video — to build Vision-Language-Action (VLA) models and World-Action Models (WAM) for robots.

VLA WAM Generalization

Task Planning with Prior Knowledge

We make robot task planning more efficient by injecting prior knowledge (e.g. commonsense knowledge) represented with structures such as ontologies — so robots reason about what to do with far less search and supervision.

Ontology Commonsense Knowledge Task Planning