Paper
PhD Candidate | EPFL
Chenghao Xu
“Anything one man can imagine, other men can make real.” — Jules Verne
I am currently pursuing my Ph.D. at École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, under the supervision of Prof. Olga Fink and in close collaboration with Schindler. Before this, I completed my M.Sc. in Robotics at Delft University of Technology (TU Delft), the Netherlands, and a B.Eng. in Mechanical Engineering with distinction from Southern University of Science and Technology, China.
During my master's studies, I worked on robust dynamic visual SLAM systems and realistic simulations of dynamic environments with Prof. Aamir Ahmad at the Max Planck Institute for Intelligent Systems, Tübingen, Germany.
Inspired by the fantasies of Jules Verne and Isaac Asimov, I am captivated by the elegance of intelligent systems, which propels me to explore the intersection between the physical world and artificial intelligence. My current research lies in 3D vision and scene reconstruction, with a particular focus on multimodal physical scene reconstruction for building assessment and renovation.
News
- Honored to be selected as one of the Outstanding Reviewers for ECCV 2026.
- Delighted to present my work at PHME 2026 and IMC 2026.
- ChatGarment is accepted to CVPR 2025.
- BuildNet3D got accepted by Building and Environment Journal.
- Excited to present BuildNet3D at FoC 2024 and AMLD 2025.
- My first paper DynaPix SLAM is accepted to DAGM GCPR 2024.
- Our research on upper-limb exoskeleton has been accepted by TMECH.
- I'm thrilled to attend the ETH Robotics Summer School this summer!
- Our work got accepted to ICRA 2023 Workshop on Active Methods in Autonomous Navigation.
- Our work got accepted to ICRA 2023 Workshop on Pretraining for Robotics.
- GRADE was accepted for presentation at NVIDIA GTC 2023.
- I will work as a research assistant at MPI-IS this summer.
- I will work on novel and impactful solutions with Spot robots in YES! Delft Impact Lab 🐕
- I am currently working as a Computer Vision R&D Engineer at Lely Technologies 🐄
- First time in Beijing: I will work as a control engineer at ROKAE Robotics.
- Admission to Master Robotics at Delft University of Technology.
- I graduated from Southern University of Science and Technology with Excellent Graduate Honor!
Selected Publications
Paper
Spline-Based Boundary Representations for Sparse View Reconstruction and Simulation Using Isogeometric Analysis
Paper
SEAR: Simple and Efficient Adaptation of Visual Geometric Transformers for RGB+Thermal 3D Reconstruction
Paper
Loc²: Interpretable Cross-View Localization via Depth-Lifted Local Feature Matching
We propose an accurate and interpretable fine-grained cross-view localization method that estimates the
3-DoF pose of a ground-level image by matching local features with a reference aerial image. The matched
ground points are lifted into BEV space with monocular depth predictions, followed by scale-aware
Procrustes alignment for pose estimation.
Paper
GRADE: Generating Realistic and Dynamic Environments for Robotics Research with Isaac Sim
In this work, we present a fully customizable framework for generating realistic animated dynamic
environments (GRADE) for robotics research. The data produced can be post-processed, e.g. to add noise,
and easily expanded with new information using the tools that we provide. To demonstrate GRADE, we
generated an indoor dynamic environment dataset and then compared different SLAM algorithms on the
produced sequences.
Paper
Exploiting Semantic Scene Reconstruction for Estimating Building Envelope Characteristics
The precise assessment of geometric building envelope characteristics is essential for informed building
retrofitting decisions. Previous methods primarily relied on detection or segmentation on 2D images.
This work leverages neural surface reconstruction based on SDF representations for 3D building analysis
and proposes BuildNet3D to estimate key building characteristics from 2D image inputs.
Paper
ChatGarment: Garment Estimation, Generation and Editing via Large Language Models
ChatGarment is a novel approach that leverages large vision-language models (VLMs) to automate the
estimation, generation, and editing of 3D garment sewing patterns from images or text descriptions.
Unlike previous methods that often lack robustness and interactive editing capabilities, ChatGarment
finetunes a VLM to produce GarmentCODE, a JSON-based, language-friendly format for 2D sewing patterns.
Paper
DynaPix SLAM: A Pixel-Based Dynamic Visual SLAM Approach
Visual Simultaneous Localization and Mapping (V-SLAM) methods achieve remarkable performance in static
environments but struggle with moving objects that affect their core modules. To address these issues,
we introduce DynaPix, a semantic-free SLAM system based on per-pixel motion probability estimation and
improved pose optimization.
Paper
Implementation of a Long-Lasting, Untethered, Lightweight, Upper Limb Exoskeleton
To prevent muscle fatigue or disorder from long-term or repetitive arm-lifting in manual operations,
various exoskeletons have been developed. To solve this problem, we designed a lightweight upper-limb
exoskeleton capable of providing self-adaptable support based on linkage mechanisms and gas springs.
Talk
Breaking the Wall of Intensive Work Above Head: Design of Passive Upper-Limb Exoskeleton
This talk presented the design of a passive adjustable upper-limb exoskeleton for intensive overhead work,
based on a spring-slider model and a four-bar linkage mechanism.
Featured Projects
Generating Realistic Animated Dynamic Environments
With GRADE framework we generate photorealistic indoor environment datasets consisting of static/dynamic scenarios and extended assets (motion blur, sensor noise, etc.). Generated data has been extensively tested on various SLAM frameworks and typical detection/segmentation libraries to prove usability and improved performance.
Lightweight Adaptive Upper-Limb Exoskeleton
The passive adjustable arm-exoskeleton is designed based on a spring slider model and four-bar-linkage model. It is a lightweight wearable system with a weight of 2 kg and with a feature of easy adjustability.
Multi-Camera Real-Time Surveillance VIDEO Stitching
Based on the AutoStitch framework, the feature matching strategy is developed given the corresponding ROIs since the cameras for surveillance are of constant parameters. Furthermore, seam-based optimization will be implemented to improve the stitching performance of the overlapping area.
Online Trajectory Planning for Manipulators Based on Discrete-Time Double-S Profile
CODE
Online Planning
Implemented real-time path following movement based on the PID method and double S profile. The constraint-based PID method can achieve synchronous movement for all joints within dynamics constraints.
TIAGo Robot for Expiring Items Picking in Retail Environment
Constructed ROS behavior tree architecture to dynamically adjusts goals and performs items picking/placing in sequence.
Machine Learning for Car Racing Games
Course
Machine Learning
Developed the Random Forest and Convolutional Neural Network models for multi-class classification, which used the current top-view image as input and outputted the control action (accelerate, steer left/right, brake).
Obstacle Detection and Avoidance for Autonomous Vehicle
Course
ROS
C++
Developed software on ROS to achieve autonomous driving in a simulated test track. Designed ROS nodes to detect obstacles and pedestrians from LiDAR pointclouds and camera images using PCL and OpenCV, and use these detections to generate simple control instructions.
Path Planner for Quadrotor Based on Kinodynamics RRT* and k-PRM Methods
Developed RRT* and k-PRM path planner to generate collision-free path to verify the robustness on 3D random obstacle map. Furthermore, vehicle routing problem will be implemented to achieve path planning for multi-rbots with multiple goals.