Robot learning · Human-in-the-loop learning · Reliable embodied AI
I'm a first-year Ph.D. student in Computer Science at UCLA, where I'm fortunate to be advised by Prof. Yuchen Cui in the Robot Intelligence Lab. My research interests lie at the intersection of robot learning, human-in-the-loop learning, and safe and reliable embodied AI. I'm particularly interested in developing robot policies that can learn from human feedback, be interpretable and steerable, and operate reliably under real-world uncertainty.
I was selected as a Fulbright Scholar by the Fulbright Commission and subsequently completed my M.S. in Electrical and Computer Engineering at UCLA. Before joining UCLA, I was a research intern at the Max Planck Institute for Intelligent Systems, where I worked with Prof. Metin Sitti on untethered magnetic milli-robots for biomedical applications. My research experiences have spanned both the physical design and control of robotic systems and the development of learning-based approaches to robot intelligence, shaping my interest in building robots that are not only capable, but also reliable and trustworthy.
SpotlightConference on Robot Learning (CoRL) 2026
Draw contact as arrows on the camera image and any image-conditioned policy, pretrained VLAs included, can use touch with no architecture change: +15.7 pp on LIBERO, +30 pp for π0.5 on real contact-rich tasks, with a sub-$50 open-source sensor.
SpotlightConference on Robot Learning (CoRL) 2026
Constrained decoding over meta-tokens guarantees that LLM robot plans satisfy syntax, preconditions, goals and ordering while keeping the LM's plan quality; a small open model reaches the best reported subgoal success on WAH-NL.
JournalNature Communications 14, 3320 · 2023
A soft magnetic robot with overlapping metal scales, like a pangolin's: it shape-morphs under low-frequency fields and heats above 70 °C from more than 5 cm away under high-frequency fields.
Adapted the ASR Conformer to decode phonemes from 256-channel intracortical recordings (Brain-to-Text ’24), adding day-specific input layers, SpecAugment and intermediate CTC. Phoneme error rate 0.161 vs. 0.217 for the GRU baseline, with no language model.
TDS convolutions for short muscle-activation patterns feeding a three-layer LSTM, plus amplitude scaling and tuned SpecAugment, on Meta’s emg2qwerty. Test character error rate 13.29 vs. 21.91 for the baseline.
Robot-centric navigation function U = γk/β: an attractive goal term over a product of obstacle barrier terms, with Kinect RGB-D colour segmentation locating the target and obstacles. Its normalized gradient steers the TurtleBot in ROS and Gazebo. The thumbnail runs the same field live.
Graduate Student Researcher · with Prof. Yuchen Cui
Asynchronous inference for diffusion and VLA policies (π0, π0.5): execution-reference anchoring lifts simulated manipulation success from 26.5% to 58.5% under 600 ms latency. Execution-supervised critics for robot program generation (AUROC 0.76 → 0.91). DARPA-funded work with Prof. Guy Van den Broeck on HMMs and Bayesian inference.
Undergraduate Researcher · with Profs. Kemal Cılız and Mehmet Akar
Built a rotary inverted pendulum and compared LQR, full-state feedback and PID on hardware and in Simulink.
Research Intern, Physical Intelligence Dept. · with Prof. Metin Sitti
Modeled the nonlinear dynamics of magnetically actuated robots to close the sim-to-real gap (→ Nature Communications); 3D tumor segmentation in brain MRI.
