Real World Reinforcement Learning of Active Perception Behaviors
NeurIPS 2025
Keywords: Active Perception, RL, VLA
I'm an incoming research scientist at Google Deepmind working on dexterous manipulation and RL for humanoids. I am also a CS Ph.D. student at the University of Pennsylvania advised by Dinesh Jayaraman. Previously, I received my BS/MS in CS at the University of Southern California, working with Joseph J. Lim on RL. I also had a wonderful time interning at Microsoft AI Frontiers with John Langford and Alex Lamb.
I find it fun to research robot learning, particularly RL and manipulation.
Click to see expanded research summary.
- World Models and Planning: World models unlock new capabilities for agents, such as solving problems with long-term dependencies and sample-efficient policy training. I've developed model-based RL algorithms that can learn sophisticated behaviors that conventional RL agents struggle to reproduce: learning to find objects in the dark, stacking blocks without rewards, and even zero-shot transfer to new robot arms. Beyond robots, I've shown that LLMs are better at planning when trained with world modeling objectives. Scaffolder, Belief State Transformer, PEG, RAC,
- Sensory Requirements of Policy Learning: Artificial agents observe the world through some input stream (e.g. vision, language, propioception). The choice of inputs greatly affects both the training dynamics and resulting behavior of the agent. My research investigates this close-knit relationship between sensing and learning, and has yielded interesting implications on the sensory requirements of RL agents, and practical RL/IL algorithms that can operate with minimal sensing. Scaffolder, AAWR, VoSI
- Learning comprehensive behaviors: For a robot to actually master a task and be useful, it must demonstrate capabilities beyond just task solving. Exploration, resetting, and verification abilities are essential for a robot to adapt to new conditions and perform tasks robustly. We've developed frameworks to teach robots these capabilities efficiently. For example, our robot learns to tug on doors after locking them to make sure they're secure, and reset their own workspaces to continue practicing. PEG, LIRF, MoreFree
Code and reviews for all of my PhD papers are public. Check them out!
Privileged Sensing Scaffolds Reinforcement Learning
ICLR 2024 (Spotlight, 5% accept rate, 3rd Highest Rated Paper in ICLR)
Keywords: Privileged Information, World Models, RL
Training Robots to Evaluate Robots: Example-Based Interactive Reward Functions for Policy Learning
CORL 2022 (Oral, 6.5% accept rate, Best Paper Award)
Keywords: Interactive Perception, Task Specification, RL
Code and reviews for all of my PhD papers are public. Check them out!
Privileged Sensing Scaffolds Reinforcement Learning
ICLR 2024 (Spotlight, 5% accept rate, 3rd Highest Rated Paper in ICLR)
Keywords: Privileged Information, World Models, RL
Training Robots to Evaluate Robots: Example-Based Interactive Reward Functions for Policy Learning
CORL 2022 (Oral, 6.5% accept rate, Best Paper Award)
Keywords: Interactive Perception, Task Specification, RL