Robot Learning · Human–Robot Collaboration
I'm a Ph.D. candidate at TTIC advised by Professor Matthew R. Walter in the Robot Intelligence through Perception Lab (RIPL). Before TTIC, I completed my master's in Computer Science at the University of Chicago and my undergraduate studies at Xiamen University.
My research focuses on robot learning and human–robot collaboration, with the goal of building robots that can adapt to unfamiliar situations and work reliably alongside people. I develop generative methods for shared autonomy and policy adaptation, using diffusion, consistency, and flow models to refine robot actions while preserving human intent or task goals. I also study visual grounding and hallucination in vision-language models, with an emphasis on understanding and detecting failures in multimodal perception.
Outside research, I enjoy photography, Chinese painting, and playing the Erhu.
GLOVES uses flow matching to selectively pull out-of-distribution actions from humans or pretrained policies toward expert behavior—without updating the underlying policy. With limited demonstrations, it improves imperfect imitation-learning and VLA agents across simulated and real-robot tasks.
SABRE turns natural-language test specifications into automatically generated, filtered, and human-verified VLM stress tests that can evolve with the models. Its SABRE-Prior benchmark exposes heavy reliance on world priors: six frontier VLMs score only 17.8–31.3% macro accuracy.
Auditing five popular robot-manipulation benchmarks with four diagnostics reveals shortcuts, statistically unresolved gains, overfitting, and training-data proximity effects—especially on LIBERO, CALVIN, and SimplerEnv. The results show why benchmark scores alone often do not establish general manipulation progress.
Across seven open VLMs, benchmark accuracy often remains surprisingly high after fine-grained visual evidence is masked, corrupted, or replaced: removing 75% of image tokens reduces POPE accuracy by only about three percentage points for some models. These results reveal a gap between benchmark performance and genuine visual grounding.
FlashBack distills diffusion into a one-step consistency-model copilot that corrects user actions in real time without knowing their goal, reward, or policy. It outperforms iterative diffusion assistance on demanding control tasks and raises real-robot peg-insertion success from 66.7% to 83.3%.
StackGen learns to generate diverse 6-DoF block arrangements that are both stable and faithful to a user-provided silhouette, avoiding explicit physics models or combinatorial search. It achieves 85.2% simulated stability and successfully builds all eight tested structures with a UR5 robot.
Partial diffusion for shared autonomy: noise and denoise a user's command toward demonstrated expert behavior, with a single knob trading off intent preservation and assistance. It needs no reward, dynamics, predefined goals, or user-policy data during training and improves both simulated and real-robot teleoperation.
HALP predicts whether a VLM will hallucinate before decoding by probing its visual and multimodal hidden states in one forward pass. Across eight VLMs, lightweight probes reach up to 0.93 AUROC while adding less than 1% overhead relative to full generation.
egoQGNN decomposes arbitrarily large graphs into ego-graphs that fit a fixed-size quantum device and uses a trainable Euclidean-to-Hilbert mapping to limit information loss, reaching competitive accuracy on six molecular and protein datasets with just 43 parameters.
Implemented Code as Policies on a UR5 robot so it could respond to voice commands and execute tasks. Presented at the 2023 National Robotics Week at MSI Chicago and the 2023 South Side Science Festival.
AI travel assistant for independent tourists from China and Japan, covering attractions, restaurants, accommodation, transportation, and language barriers.
