Hengxiang Chen

A Generalizable Approach for Safety-Aware Robotic Behavior Planner with Constraint Space

Xianxu Wu, Shang-Ching Liu, Hengxiang Chen, Qiang Li, and Jianwei Zhang

CASE2026 In review

paper project

Octopi-X: Robotic Perception with a Large Tactile-Vision-Language Model for Physical Property Inference

Zexiang Guo*, Hengxiang Chen*, Xinheng Mai*, Qiusang Qiu, Gan Ma, Zhanat Kappassov, Qiang Li, Nutan Chen

Best Paper Finalists Award in IROS2025 Workshop Tactile Sensing

paper project arxiv

Inferring physical properties can significantly enhance robotic manipulation by enabling robots to handle objects safely and efficiently through adaptive grasping strategies. We integrate visual observations with tactile representations within a multimodal vision-language model and show strong zero-shot generalization.

Cross Modal Robotic Perception with a Large Vision Language Model for Physical Property Inference

Zexiang Guo*, Hengxiang Chen*, Qiusang Qiu, et al.

CLAWAR2025

paper

Inferring physical properties can significantly enhance robotic manipulation by enabling robots to handle objects safely and efficiently through adaptive grasping strategies. Our framework fuses visual observations and tactile representations, outperforming baselines on diverse objects.

Path Planning Algorithm Comparison Analysis for Wireless AUVs Energy Sharing System

Zhengji Feng, Hengxiang Chen, Liqun Chen, Xiaolin Mou

IEEE ITeN 2023

paper code

This research proposes a wireless AUV energy sharing system for underwater recharging and compares RRT* and PSO for path planning under static and dynamic obstacles, with analysis of path length, energy consumption, and computational complexity.

本科毕业论文：移动操作机器人定位算法和基于视触融合的抓取研究
Graduated Thesis: Vision-Touch Guided Grasping for Mobile Manipulators control of remote urban driving

Hengxiang Chen, Qiang Li

code

Problem: Mobile manipulators often struggle with accurate object grasping and long-horizon navigation in dynamic environments.
In this work, I designed a full-stack robotic software system based ROS combining 'Cartographer' for high-precision localization, 'GraspNet' and 'YOLOv8' for object detection and 6-DoF grasping pose generalization, and a close-loop control algorithm using tactile optical sensors(Gelsight Mini), trying to let the mobile manipulator deliver books to humans even in a cross-floor scenario.

Internship Technical Report: The impact of latency and vehicle speed on the motion control of remote urban driving

Hengxiang Chen, Yongwei Yang

In this research, we quantitatively analyzes the impacts of latency and vehicle speed on motion control in remote urban driving through statistical modeling of simulation results and real-world vehicle data. It employs Time To Collision (TTC) as a collision risk metric to evaluate and quantify the risk associated with different latency scenarios. Experimental results indicate that higher latency significantly increases collision risks in specific traffic scenarios, such as protected left turns and intersections. The study proposes practical safety countermeasures, including scenario restrictions, speed optimization, and local Automated Emergency Braking (AEB), providing critical insights to improve the safety and reliability of remote driving systems.

Chinese Robotics and Artificial Intelligence Competition (Intelligent Driving)

Team Leader

June 2023 - Hainan, China

5th Place (National First Prize)

In this competition, we aimed to reach the goal area as quickly as possible on a predefined map consisting of straight roads, curved racing sections, and a figure-eight roundabout.
I was mainly responsible for the development of system pipeline based on ROS, I also participated in fine-tuning parameters of 'Movebase', which included many autonomous algorithms for mobile base planning and control.

Chinese Outdoor ROS Autonomous Racing Competition

Team Leader

December 2022 - Shenzhen, China

3rd Place (National First Prize)

The objective of this competition was to design and implement the software system for an autonomous race car capable of completing two laps of a cone-defined track in the shortest possible time.
In this work, I developed the method of multi-sensors fusion as following.