Ran Ding

I am an MSc student in Informatics at the Technical University of Munich and a research assistant at the TUM Computer Vision Group, advised by Dr. Yan Xia , and supervised by Prof. Daniel Cremers . I also collaborate closely with Prof. Hao Dong at the PKU Agibot Lab. Previously, I earned a BEng in Computer Science from Dalian University of Technology.

I have a strong mission to make robots intelligent. This interest ignited during my undergraduate studies, where I worked on robotics under the supervision of Prof. Zhenyu Wu. Through this experience, I discovered my immense passion for robotics and became aware of the limitations of the traditional methods.

My current research focuses on embodied AI, computer vision, and multimodal learning. Specifically, I am currently working on perception for robot grasping and manipulation.

Email  /  GitHub  /  LinkedIn  /  Twitter

Research

TARGO: Benchmarking Target-driven Object Grasping under Occlusions

This paper presents TARGO, a new benchmark dataset focused on target-driven grasping under occlusions in cluttered environments. It investigates how occlusion affects grasping, revealing that even state-of-the-art models struggle as occlusion levels increase. To address this, TARGO-Net, a transformer-based model with a shape completion module, was developed to improve grasping performance in these challenging conditions. Topics: Target-driven Robot grasping, Occlusion Handling

Deformation Generation via Autoregressive Models

Given partial deformation, we applied a decoder-only transformer to autoregressively generate diversity and complete deformation. Topics: Deformation Generation, 4D Completion, Autoregressive Model

Projects

It covers computer vision, robotics, and multimodal learning.

Unsupervised Multi-View Stereo with Neural Radiance Field

We extended the Neural Radiance Field in the field of Multi-View Stereo and trained it in an unsupervised manner to resolve such ambiguity issues of view correspondences. Topics: Multiview-Stereo, Neural Radiance Field, Unsupervised Learning

Stereo Reconstruction

We implemented a traditional stereo reconstruction pipeline consisting of SfM and MVS and analyzed the qualitative and quantitative performance of various approaches. Topics: Traditional 3D Reconstruction, Structure from Motion

National Undergraduate Electronic Design Contest

We developed a target tracking system that controls a vehicle to follow a predefined path and implements algorithms for data communication and parsing between the millimeter-wave radar and the vehicle. Topics: MCU Development, Radar sensor, PID

Chinese Undergraduate Curling AI Challenge

We refined motion strategies for curling robot, including curve shots/hits and implemented a neural network regression model to accurately predict the relationship between velocity (both linear and angular) and the current position of the robot. Topics: Robot decision, Neural network regression

National Undergraduate Electronic Design Contest Liaoning Division

We developed a paper counting device, utilizing the least squares method to establish a correlation between paper quantity and capacitance data collected by the FDC2214 capacitance sensor, while enhancing the communication among the various modules. Topics: MCU Development, Least squares method, Capacitance sensor

Awards