Embodied AI · Robot Learning · Imitation Learning
M.S. in Electrical & Computer Engineering,
University of Southern California.
MD
Hi there! I'm an M.S. student in Electrical & Computer Engineering at the University of Southern California. I build robots that learn efficiently and reliably — from expert demonstrations, reward signals and autonomous interaction.
I work with Jesse Thomason in the GLAMOR Lab on composing vision-language-action policies during execution, and with Erdem Bıyık in the LIRA Lab on demonstration-efficient imitation learning. Currently I am interning at the Georgia Tech Research Institute with Colin Usher on vision-based pick-selection for deformable objects.
I'm most excited about where this goes next: robots that adapt on their own, from interaction rather than new supervision.
DiPS, our work on dialogue policy selection for high-stakes persuasion agents, was accepted to SIGDIAL 2026!
Submitted SWAP, a stepwise action policy-routing framework for VLAs, to CoRL 2026.
Started as a Research Intern at the Georgia Tech Research Institute with Colin Usher, working on robotic pick-selection for deformable objects.
Joined the LIRA Lab at USC with Erdem Bıyık, working on teacher-aware imitation learning.
Joined the GLAMOR Lab at USC with Jesse Thomason, working on language-guided robots.
Started my M.S. in Electrical & Computer Engineering at the University of Southern California.
It is hard to grasp multiple densely-packed deformable objects in clutter. We learn a pick ranker from operator demonstrations that continually improves from human corrections at deployment, on top of perception and grasping foundation models.
In preparation. Target: ICRA 2027.
Human demonstrators don't act randomly — they teach. TV-BC upweights the hard, informative demonstrations a teacher would emphasize, learning more sample-efficiently and generalizing better out-of-distribution than standard behavioral cloning.
In preparation. Target: ICLR 2027.
No single policy is best at every step of a task. SWAP learns to switch between vision-language-action policies mid-rollout, improving real-robot success by up to 33% over the best fixed policy.
Submitted to CoRL 2026.
Persuasion isn't one-size-fits-all. In a wildfire-evacuation setting, DiPS learns to pick the right persuasion strategy turn-by-turn, outperforming zero-shot and RAG baselines with both simulated and real people.
SIGDIAL 2026 (Oral Presentation)
A hybrid DenseNet201 + Mediapipe model for robust hand-gesture recognition.
ICACRS, IEEE 2022.
* denotes equal contribution.
Pushing a leg harder into sand doesn't always climb faster. On a custom 2-DoF-per-leg quadruped and an adjustable sand ramp, we swept stride amplitude against slope to find a sweet spot where moderate strides gain traction without avalanching the substrate, tripling uphill displacement over the baseline.
A robotic hand that mirrors a user in real time using two input modes at once: speech recognition for voice commands and Mediapipe landmark tracking for live gesture replication. Detected finger-joint angles map to servo commands on a microcontroller, keeping end-to-end latency under 100 ms.
A swarm of mobile robots sort packages under an overhead camera, each one tracked by its own ArUco marker. A central planner assigns A* routes and relays them over WiFi to onboard microcontrollers, cutting sort-cycle time by 35% against a fixed-route baseline.
I'm always happy to discuss robot learning and research collaborations.