Computer Science PhD, Stanford University
I am a PhD student in the Department of Computer Science at Stanford University, advised by Chelsea Finn and Dorsa Sadigh. I am currently also working at Google DeepMind. My research focuses on reinforcement learning. Prior to Stanford, I received my B.S. and M.S. from UC Berkeley, working with Sergey Levine and Yi Ma.
arXiv preprint arXiv:2609.18207, 2026
Presents a framework for reinforcement learning fine-tuning of real-time policies through decoupled slow action generation and fast, reactive editing, combining the reliability of RL with the reactiveness required for dynamic control.
arXiv preprint arXiv:2607.27203, 2026
Shows that standard Q-function pretraining gives little benefit over random initialization, and introduces Initialization via Policy Ensemble (IPE), which bootstraps Q-learning from the pooled rollouts of several diverse pretrained policies for more effective online RL fine-tuning.
Conference on Robot Learning (CoRL), 2026
Introduces EXPO-FT, a system for reliable, sample-efficient RL finetuning of pretrained vision-language-action models that reaches perfect task success across all evaluated manipulation tasks using an average of just 19.1 minutes of online robot data.
arXiv preprint arXiv:2604.19730, 2026
Presents FASTER, which models action-candidate denoising as an MDP and learns a value function in denoising space to filter candidates, recovering the benefits of test-time sampling in RL without its computational overhead.
International Conference on Learning Representations (ICLR), 2026
Uses flow matching to model the full continuous distribution over future returns in distributional RL, avoiding the discretization into bins or fixed quantiles required by prior categorical or quantile-based approaches.
International Conference on Learning Representations (ICLR), 2026
EXPO is a highly sample-efficient online RL algorithm for expressive policies that pairs a large expressive base policy trained via a stable imitation objective with a lightweight edit policy that shifts sampled actions toward higher-value regions for stable value maximization.
International Conference on Learning Representations (ICLR), 2026
Conducts a systematic empirical study of the factors that make batch online RL — policy improvement from large batches of autonomously collected robot data — effective, addressing why prior imitation-learning-based approaches often fail to improve.
International Conference on Machine Learning (ICML), 2026
Identifies attention-score collapse as the key failure mode preventing transformer value functions from scaling in RL, and proposes Transformer Q-Learning (TQL), which controls attention entropy to stabilize training as network size grows.
arXiv preprint arXiv:2512.16911, 2025
Shows that standard behavioral cloning can fail to cover the demonstrator's action distribution, and proposes Posterior Behavioral Cloning (PostBC), which models the posterior over demonstrator behavior to ensure coverage while matching BC's pretrained performance.
arXiv preprint arXiv:2506.07505, 2025
Introduces Data-Guided Noise (DGN), a sample-efficient RL method that uses prior demonstration data to guide exploration noise rather than explicitly cloning demonstrated actions, improving over prior RL-from-offline-data methods across continuous control tasks.
International Conference on Learning Representations (ICLR), 2024
Proposes treating human intervention signals themselves as the RL reward in interactive imitation learning, in place of methods like DAgger that assume a near-optimal intervening expert, enabling policies that can surpass a suboptimal expert.