Rongxuan Zhou

About

I am currently a graduate research assistant at the CRAFT Lab, Northeastern University, advised by Prof. Gilbert Yang Ye. My research focuses on safe robot learning and manipulation.

Before joining CRAFT Lab, I worked as a research assistant at SJTU IWIN-FINS Lab advised by Prof. Jianping He, where I co-designed the full-stack Fines robotic platform, covering hardware architecture, embedded control framework, and perception pipeline.

On the industry side, I worked as an automation engineer and designed production AOI (Automated Optical Inspection) equipment for Texas Instruments semiconductor products at Dinnar Automation, co-designed standardized multi-robot assembly workstations for harmonic reducer production with EtherCAT real-time control and digital twin systems at CloudMinds Robotics, and developed sim-to-real transfer pipelines for autonomous driving at NIO Automobile.

Download CV

Publications

Deploying...

Is Your Policy Actually Using Force? Why Concatenation Fails for Contact-Force Integration in Sequential Robot Policies

Rongxuan Zhou and Gilbert Yang Ye
In Preparation Target: CoRL 2026

[Abstract] [Project] [Code]

Multi-modal sequential policies fuse proprioception, vision, and auxiliary modalities by concatenating them into a shared processing sequence — yet on a contact-rich nut-threading task, naïve raw-force concatenation reduces success rate by 32 percentage points compared to ignoring force entirely. We show that FiLM side-channel conditioning bypasses competitive normalization in self-attention and LSTM gating layers, restoring force utility (+17 pp on NutThread, non-overlapping 95% CIs). To diagnose suppression post-hoc without retraining, we introduce R_contact, a contact-sensitivity ratio via Integrated Gradients that reliably distinguishes suppressed (≈1×) from functional (3.0–3.7×) architectures. Validated across 5 contact-rich tasks on two simulators (robosuite imitation learning + Isaac Lab Factory reinforcement learning).

TACS: Training-Free Safety Guidance for Action-Space Diffusion Policies via Tweedie-Adaptive Constrained Sampling

Rongxuan Zhou and Gilbert Yang Ye
Under Review IEEE Robotics and Automation Letters (RA-L)

[Abstract] [Project] [Code]

Diffusion policies have become the dominant paradigm for learned robotic manipulation, yet they offer no intrinsic mechanism to enforce workspace safety constraints. TACS is a training-free method that injects barrier-function guidance into the DDIM denoising loop of action-space diffusion policies, applying repulsive guidance on the Tweedie-predicted end-effector positions; subsequent denoising steps coherently adapt rotation and gripper dimensions, achieving effective obstacle avoidance at 10× smaller guidance scales than post-hoc potential-field projection. On three robosuite manipulation tasks, TACS Pareto-dominates post-hoc filtering, with a statistically significant win-win on Lift (violations 6.6% → 0.32%, 20× reduction; p=0.015 across 5 seeds × 250 episodes) and zero measurable inference overhead versus the unmodified diffusion policy.

A Robust Distributed Odometry for Mobile Robots with Steerable Wheels

W. Xi, J. Guo, C. Wang, S. Wu, and J. He
IROS 2025 Hangzhou, China
Contribution: Lead designer of mechanical architecture and simulation environment for the “Fines” steerable wheeled mobile robot platform

[Abstract] [Project] [Code]

This paper presents a robust distributed odometry framework for mobile robots equipped with steerable wheels, fusing wheel encoder and IMU measurements through a factor-graph optimization backend for accurate real-time localization.

Simulation Research on Residual Stress of Swage Autofrettage-processed High-Pressure Cylinder

L. Sun, R. Zhou, G. Li, J. Li, and P. Mitrouchev
Published J. Phys.: Conf. Ser., vol. 2587, 012088, 2023. (EI Indexed)

[Abstract] [Project] [BibTeX]

This study investigates the residual stress distribution in swage autofrettage-processed high-pressure cylinders through finite element simulation, analyzing the effects of mandrel interference percentage on von Mises and circumferential residual stresses to optimize cylinder fatigue life.

@article{sun2023simulation,
  title={Simulation Research on Residual Stress of Swage
         Autofrettage-processed High-Pressure Cylinder},
  author={Sun, L. and Zhou, R. and Li, G. and Li, J. and Mitrouchev, P.},
  journal={J. Phys.: Conf. Ser.},
  volume={2587},
  pages={012088},
  year={2023}
}

Selected Projects

Automated Optical Inspection System for Texas Instrument

Sole engineer on 4-CCD multi-angle visual inspection system for TI semiconductor CSE products. 19 defect categories, 100% detection rate, >85K units/day production throughput. Full mechanical design of 1800×1600×2000 mm equipment with 18-step automated process flow.
Machine Vision AOI CAD ANSYS Semiconductor
85K+ units/day 100% detection 19 defect types

[Project] [Code]

Designed and delivered a production AOI (Automated Optical Inspection) system for Texas Instruments CSE semiconductor products at Dinnar Automation. The system uses 4 industrial CCD cameras (MV-GE501GC) with multi-angle illumination to inspect 19 defect categories including epoxy exposure, pin missing, contamination, and marking defects, achieving 100% detection rate at >85K units/day throughput.

EtherCAT Real-Time Manipulator Control

Developed EtherCAT master module with 1 kHz F/T sensor synchronization, reducing joint control cycle from 20 ms to 12 ms (40%). Deployed data-driven impedance controller on UR5e with 42% vibration reduction, validated through 72-hour aging test.
C++ EtherCAT CODESYS Real-Time UR5e

[Project] [Code]

Developed EtherCAT master module with 1 kHz F/T sensor synchronization at CloudMinds Robotics, reducing joint control cycle from 20 ms to 12 ms (40%). Co-designed standardized multi-robot assembly workstations for harmonic reducer production with digital twin systems. Deployed data-driven impedance controller on UR5e with 42% vibration reduction.

Fines Full-Stack Robotic Platform

Co-designed hardware architecture, embedded control framework (FineMote), and perception pipeline (FineVision) for steerable wheeled mobile robot at SJTU. Platform enabled experimental validation for IROS 2025 distributed odometry paper.
C++ ROS2 LiDAR Embedded MoveIt!
1 ms control cycle IROS 2025 validated

[Project] [Code]

Co-designed the full-stack "Fines" robotic platform at SJTU IWIN-FINS Lab: hardware architecture (steerable wheeled chassis), embedded control framework (FineMote, 1 ms cycle), and perception pipeline (FineVision with LiDAR + IMU fusion). The platform enabled experimental validation for the IROS 2025 paper on distributed odometry.

Autonomous Driving Sim-to-Real Transfer

Built GAN-based CAN FD signal synthesizer (5× augmented data, +35% ECU test coverage) and gradient reversal domain adaptation (CARLA→NIO ET5) reducing lateral control error by 37% in production vehicle road tests.
PyTorch GAN CARLA Sim-to-Real
37% error reduction Production validated

[Project] [Code]

Built GAN-based CAN FD signal synthesizer (5× augmented data, +35% ECU test coverage) and gradient reversal domain adaptation pipeline (CARLA→NIO ET5) at NIO Autonomous Driving, reducing lateral control error by 37% in production vehicle road tests.

MuJoCo MCP Server — AI-Driven Robotics Simulation

65-tool server connecting MuJoCo physics simulation to AI assistants via natural language. Supports trajectory optimization (iLQR, MPPI), inverse kinematics, domain randomization, and RL environment integration for 50+ robot models from MuJoCo Menagerie.
Python MuJoCo MCP Robotics Sim RL

[Project] [Code]

Open-source 65-tool MCP server bridging MuJoCo physics simulation with AI assistants via natural language. Supports trajectory optimization (iLQR, MPPI), inverse kinematics, domain randomization, and RL environment integration for 50+ robot models from MuJoCo Menagerie.

LiDAR-Inertial SLAM Benchmarking

Benchmarked FAST-LIO vs. LIO-SAM on real-world driving datasets collected with Northeastern’s NUance platform. Integrated FAST-LIO-LC for loop closure drift correction over long trajectories.
C++ ROS SLAM LiDAR

[Project] [Code]

Benchmarked FAST-LIO vs. LIO-SAM on real-world driving datasets collected with Northeastern's NUance autonomous vehicle platform. Integrated FAST-LIO-LC for loop closure drift correction over long trajectories, achieving consistent sub-meter accuracy.