Xi Chen

Hi! I’m Xi Chen (陈希), a second-year M.S.E. student in Computer Science at Johns Hopkins University. I am fortunate to be advised by Prof. Laixi Shi! Before JHU, I received my B.Eng. in Computer Science from Nankai University.

I want to build decision-making agents that keep working when deployment differs from training, and I am applying for a Ph.D. to pursue this direction. My research focuses on:

• 🛡️ Robust reinforcement learning — Policies that keep working at deployment: algorithms robust to dynamics shift and other perturbations (observation, action, reward, latency, semantics), and benchmarks that measure such robustness fairly.
• 🤖 Robot learning — Training policies for robotics and humanoid tasks in simulation and closing the sim-to-real gap.
• 🌍 Generative world models for RL — Generative models that imagine multi-step futures, used as data augmentation to train policies that are robust to shifted dynamics.

Earlier work spans time-series learning, generative modeling and computer vision. I enjoy problems where robust decision-making meets real deployment constraints.

---

news

selected research projects

A Unified Library and Benchmark for Robust Reinforcement Learning Algorithms   · 2026 – Present

Johns Hopkins University — with Prof. Laixi Shi and collaborators

An algorithm-centric library and benchmark for robust RL under deployment uncertainty. It integrates 22 algorithms (6 standard, 16 robust online, offline and safe), organizes them by mechanism and claimed uncertainty, and evaluates every frozen policy under one protocol on a shift toolbox of six sources (dynamics, observation, action, reward/cost, latency, semantic) with several modes and their compositions, across hundreds of perturbation configurations. The same interface runs on advanced robotic tasks — humanoid and manipulation in Isaac Lab and a frozen vision–language–action policy — to mimic the sim-to-real gap. The findings: robustness is channel-specific, and what a mechanism models is what it defends. Under review. Library and project website to be released after the review period.

Generative Uncertainty Modeling for Distributionally Robust Offline RL   · Apr 2026 – Present

Johns Hopkins University — with Prof. Laixi Shi and collaborators

A robust offline-RL framework that builds the uncertainty set from a conditional flow model instead of a rectangular, divergence-ball set: shared flow parameters couple shifts across state–action pairs, and the flow-matching loss serves as the discrepancy measure, which provably implies a Wasserstein-2 bound. The flow model is adversarially fine-tuned against the current policy with an inverse-propensity-weighted, PPO-style objective under the flow-matching constraint, and the generated adverse H-step trajectories become conservative critic targets for IQL, CQL or BCQ backbones. On MuJoCo dynamics shifts and an EPANET water-distribution-network control case study it improves OOD scores without sacrificing nominal performance. Under review.

Point Tracking for Bronchoscopy via Adjacent-Frame Correlation   · Feb 2026 – Present

ARCADE Lab, Johns Hopkins University

An AllTracker-style tracker adapted to bronchoscopy, where textureless airway walls, illumination shifts, and specular reflections break long-range query-anchored matching. An adjacent-frame correlation module estimates and chains frame-to-frame optical flow for robust tracking, distilled (Transformer teacher → lightweight CNN student with D2-Net-style features) for real-time deployment at < 30 ms latency. Manuscript in preparation.

Backdoor Attacks on Multivariate Time-Series Forecasting   · Dec 2024 – May 2025

Undergraduate thesis, DBIS Lab, Nankai University — with Dr. Xiangrui Cai

A stealthy backdoor-attack framework for multivariate forecasting under realistic missing-value scenarios: invisible triggers at missing-value positions plus a critical-time-step imputation strategy preserve stealth while cutting attack MAE by 43% versus baselines.