Yixuan Yang

Research Experiences

You can also check my Google Scholar profile.

• arXiv
  Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories

  Yixuan Yang, Mehak Arora, Ryan Zhang, Baraa Abed, Junseob Kim, Tilendra Choudhary, Md Hassanuzzaman, Kevin Zhu, Ayman Ali, Chengkun Yang, Alasdair Edward Gent, Victor Moas, Rishikesan Kamaleswaran

  Preprint: arXiv:2605.10840 (May 2026) · under review
  

  [Paper] [Code]

  Every JEPA before this one either discards the predictor or trains it on a frozen encoder, so the encoder never learns to support the rollout it will be asked for. Clin-JEPA co-trains them, and the representation spreads out: 55% more effective dimensions, variance no longer piled into a handful of them. Freeze that same encoder, train a predictor the ordinary way, and long-horizon accuracy falls below doing nothing.

• 
  Contact-Rich Humanoid Manipulation for Scientific Laboratory Automation

  Yixuan Yang. Advised by Professor Boyuan Chen

  General Robotics Lab, Duke University (Aug. 2024 – Mar. 2025)
  

  Demos: [Heuristic] [RL policy] [Knob 1] [Knob 2] [Knob 3] [Learning kit]

  Laboratory instruments are built for human hands: knobs, switches, small buttons. This project put a Unitree G1 humanoid in front of one, pairing a physical testbed with a dimension-matched replica in Isaac Gym. Most of the work was the task stack itself: an inverse-kinematics control interface, fingertip contact sensing, and the reward and termination design that gets PPO to learn contact-rich motion rather than flail near it.

• ACM TOSN
  SniffySquad: Patchiness-Aware Gas Source Localization with Multi-Robot Collaboration

  Yuhan Cheng*, Xuecheng Chen*, Yixuan Yang, Haoyang Wang, Jingao Xu, Chaopeng Hong, Susu Xu, Xiao-Ping Zhang, Yunhao Liu, Xinlei Chen

  ACM Transactions on Sensor Networks (Feb 2026) (DOI: 10.1145/3786599)
  

  [Paper]

  Turbulence tears a gas plume into patches, and every patch can look like the source. So SniffySquad stops climbing the gradient and starts sampling: each robot is a Langevin chain on a shared belief map, and its temperature is its role. Hot robots roam, cold robots inspect, and the two trade places by the same acceptance rule parallel tempering uses. Success rate up 20%, path efficiency up 30%.

• HotMobile
  Olfactory Sensing in Turbulent Airflow via Collaborative Robots

  Yuhan Cheng*, Xuecheng Chen*, Yixuan Yang, Haoyang Wang, Yuxuan Liu, Xinlei Chen

  HotMobile 2024: Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications
  

  [Paper]

  Turbulence scatters a gas plume into patches, so following the concentration gradient walks a robot into a decoy. This early poster gives a robot team heterogeneous roles: some inspect the strongest reading found so far, others keep hunting for new candidates, and the roles are exchanged as the readings change. Search time fell 37% against Surge-Cast and 22% against Infotaxis, which gets trapped in exactly those patches.

• IEEE CEC
  Learning-Based Problem Reduction for Large-Scale Uncapacitated Facility Location Problems

  Shuaixiang Zhang, Yixuan Yang, Hao Tong, Xin Yao

  CEC 2024: IEEE Congress on Evolutionary Computation
  

  [Paper]

  Facility location at scale defeats search heuristics, so this shrinks the problem before solving it. The trick is to look at ranks rather than costs: each facility is described by the shape of its rank distribution across customers, so the same four numbers describe it whether the instance holds fifty sites or five thousand. That scale-free footing is why a model trained on instances a solver can crack exactly transfers to ones it cannot, cutting benchmarks to 13% of their size.

• 
  Learning to Say No: Unsolvable Robotic Task Detection Using Synthetic Data Driven by LLaVA

  Final project, Duke ECE 590 / ME 555: Robot Learning (Fall 2024)
  

  [Paper] [Github]

  Robots asked to do something impossible tend to try anyway. This builds a synthetic dataset of five ways a task can be unsolvable, from a missing item to an ethical refusal, then fine-tunes a vision-language model to spot them and say why. Refusal accuracy goes from 10% to 78% on the generated images, and 81% in Habitat-Sim, a renderer it never trained on.