Research Interests: I am broadly interested in developing principled and practical pipelines that extend machines’ capability to learn in realistic problem settings. I study how machines can interpret complex systems through generative modeling to solve real-world tasks related to perception, control, and scientific discovery. I investigate ways to zoom in/out data through multi-resolution representation to capture the quantity of interest with efficiency and accuracy through techniques such as structural modeling, abstraction, and coarse-graining.
- New Preprint Crystal Diffusion Variational Autoencoder for Periodic Material Generation addressing the generation of periodic material structures with physical inductive biases using a diffusion model decoder.
- New Preprint Fragment-based Sequential Translation for Molecular Optimization that optimizes and discovers diverse molecules using an RL approach with action abstraction.
- New CoRL paper Learning to Jump from Pixels on using hierarchical control for agile locomotion with visual input on the MIT mini-cheetah robot.
- New ICML paper Learning Task Informed Abstractions on enabling model-based agents to learn better policies in cluttered visual domains.