Selected research papers (* denotes equal contribution)
arXiv preprint, 2026
Humanoid autonomy trained in procedurally generated simulated homes and deployed in the physical world. HALO was tested in six real-world environments that were neither seen during training nor scanned or...
Foundations and Trends in Robotics, 2026
Motion prediction powers everything from self-driving cars to collaborative robots—but why do models that ace benchmarks often struggle in the real world? This comprehensive survey digs into that gap. We...
IEEE International Conference on Robotics and Automation (ICRA), 2026
We propose a data-driven framework that learns a Koopman embedding for nonlinear robot dynamics, integrating the resulting linear model with the Safe Set Algorithm to solve tracking and safety constraints...
arXiv preprint, 2026
We study how the prediction and control errors of neural Koopman models scale with latent dimension and training data across 8 nonlinear systems. We establish a scaling law showing the...
Doctoral dissertation, 2025
How do we build robots that keep learning throughout their lifetime—without forgetting what they already know? My doctoral thesis tackles this grand challenge head-on. Real-world robots face streaming data in...
Transactions on Machine Learning Research (TMLR), 2025
Diffusion models generate stunning images, but what if we could make them even better? We discovered that the standard denoising process has a blind spot: it doesn’t know how “noisy”...
Annual Learning for Dynamics & Control Conference (L4DC), 2025
Teaching robots to walk is hard—keeping them walking well as terrain changes is harder. We introduce a continual learning framework that lets legged robots refine their dynamics model on the...
Conference on Parsimony and Learning (CPAL), 2025
Adam is everywhere in deep learning, but its default initialization has a subtle flaw that hurts generalization. We traced the problem to the second-order moment estimate starting at zero—a seemingly...
Arxiv preprint, 2024
Imagine a robot assistant that sees what you’re doing, listens to what you say, and seamlessly helps over extended periods. That’s what we built. Our hierarchical framework fuses vision and...
Conference on Robot Learning (CoRL), 2024
How can a robot manipulate objects it’s never seen before, using only camera images? KOROL (Koopman Operator Rollout) learns to extract object features from images and propagate them forward in...
CoRL Workshop on Safe and Robust Robot Learning for Operation in the Real World [Oral Presentation], 2024
A self-driving car’s neural network works flawlessly in testing—then fails on a slightly different road sign. Why? Neural networks can be surprisingly fragile to tiny input changes. We show that...
MedRxiv preprint, 2024
Can GPT-4 help cardiologists make better decisions? We put it to the test in cardiac electrophysiology—a complex specialty where AI assistance could really matter. Using Retrieval-Augmented Generation (RAG), we boosted...
Transactions on Machine Learning Research (TMLR), 2024
Certified training promises neural networks that are provably robust—but at what cost? We found that existing methods over-regularize, crippling the model’s performance. The culprit: unstable neurons with high variance. Borrowing...
Conference on Robot Learning (CoRL), 2023
Robots deployed in the real world encounter environments that slowly drift over time—lighting changes, surfaces wear down, sensors degrade. Standard models trained offline can’t keep up. Our solution: online feedforward...
IEEE Transactions on Robotics (T-RO), in adjunct with ICRA, 2023
Humans navigate familiar places effortlessly while still learning new ones—why can’t robots? BioSLAM brings this capability to visual place recognition through a brain-inspired dual-memory system. Dynamic memory rapidly encodes new...
IEEE Conference on Control Technology and Applications (CCTA), 2023
Most robots wait for commands—ours anticipates. We built a proactive human-robot collaboration system that predicts your intentions and acts accordingly, while safety controls ensure the robot never endangers its human...
Pattern Recognition, 2023
Pre-trained models are powerful, but fine-tuning them with limited labeled data is tricky—you risk either forgetting what they knew or overfitting to your small dataset. We introduce hierarchical self-regularization: the...
Transactions on Machine Learning Research (TMLR), in adjunct with ICML-AI4S Workshop, 2023
Can reinforcement learning master the chaos of optical systems? We built a simulation environment to find out. Real optical control is brutal—nonconvex, nonlinear, and plagued by time-dependent noise—yet traditional controllers...
ArXiv preprint, 2022
How do you build robots that can localize themselves for years, not just hours? This survey maps the landscape of long-term place recognition—the algorithms, the challenges, and the opportunities. Whether...
Advances in Neural Information Processing Systems (NeurIPS) [Spotlight], in adjunct with IJCAI-WSRL Workshop [Best Paper Runner-ups] , 2021
Drug discovery relies on predicting molecular properties—but what if you only have a handful of examples? Traditional methods fail here. Our Property-Aware Relation (PAR) Networks learn to adapt molecular relationships...
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021
What happens when you combine the best of two worlds—pre-trained models from one domain and a mix of labeled and unlabeled data from another? We developed Adaptive Consistency Regularization to...
CVPR Workshop on Learning from Limited and Imperfect Data [Invited Talk], 2021
Everyone assumed Vision Transformers need massive datasets—we proved them wrong. By rethinking model size and tokenization, our Compact Convolutional Transformer (CCT) achieves 98% accuracy on CIFAR-10 using just 3.7M parameters....
International Journal of Adaptive Control and Signal Processing (ACSP), 2021
Predicting what a human will do next is crucial for safe human-robot collaboration—but people are unpredictable, and models trained offline quickly become stale. We built an online adaptation framework combining...
Learning for Dynamics and Control (L4DC), 2020
Neural networks often fail when test conditions differ from training—the classic train-test gap. Rather than accept this, we asked: what if the model could adapt itself at test time? Inspired...
AAAI Fall Symposium Series, AI for HRI, 2019
For robots to work safely alongside humans, they need to predict what we’ll do next. We developed a multi-task model that simultaneously forecasts human trajectories and intentions—and crucially, adapts quickly...
ACL Student Research Workshop, 2019
Social media comments are deeply personal—can AI capture that? We developed a Personalized Comment Generation Network that produces relevant, grammatically correct comments matching individual personality traits and writing styles. Moving...
Work in optics and other areas
Optics Express, 2021
Coherent pulse stacking can boost laser power dramatically, but controlling it is a nightmare—the system is nonlinear, noisy, and hard to optimize. We combined the...
Optics Express, 2020
Synchronizing clocks over long distances with sub-picosecond precision sounds impossible—until you use optical frequency combs. We demonstrated linear optical sampling with dual combs over a...
Conference on Lasers and Electro-Optics (CLEO), 2020
High-power lasers require combining many weak pulses into one powerful burst—but maintaining coherence is fiendishly difficult. We achieved the first-ever coherent stacking of 128 ultrashort...