BEng student in Mechatronics and Robotic Systems at XJTLU
Working on Vision-and-Language Navigation, embodied AI, and autonomous mobile robots

ETPNav × MID360 · NaVIDA × Habitat · ROS 2 FishBot · Repositories

I am an undergraduate student at Xi'an Jiaotong-Liverpool University (XJTLU), pursuing a BEng in Mechatronics and Robotic Systems (expected 2028).

My current work sits at the intersection of robot navigation, embodied AI, and robot perception. I am particularly interested in how learned navigation policies can be connected to geometric sensing and deployed on practical robotic systems.

Research interests

• Vision-and-Language Navigation (VLN)
• Embodied AI and robot navigation
• LiDAR / RGB-D perception for navigation
• Learning-based waypoint proposal and topological planning
• ROS 2, Nav2, simulation, and robot deployment

I am studying whether an RGB-D-based VLN system can retain its learned navigation prior when dense depth is replaced by MID360-like sparse LiDAR geometry.

The project currently focuses on the waypoint interface while keeping the downstream GraphMap and language/topological planning pipeline frozen whenever possible. Work so far includes explicit-geometry waypoint baselines, proposal–GraphMap interface analysis, sparse-depth zero-shot experiments, activation-level failure forensics, and feature-domain adaptation studies.

The longer-term path is Habitat / R2R-CE → Isaac Sim native LiDAR → Livox MID360 + real robot navigation.

Public research materials: Research Showcase · Application Summary · Detailed Overview

A paper-oriented integration of the official NaVIDA checkpoint with Habitat. The project implements a stateful inference service, historical RGB observation handling, strict action-chunk parsing, and closed-loop Habitat execution.

Highlights:

• official NaVIDA-3B checkpoint running in 4-bit on an RTX 4060 Ti 8GB;
• paper-pure RGB + language inference protocol;
• stateful localhost HTTP service and episode lifecycle;
• end-to-end model → action parser → env.step() smoke testing;
• auditable logging for actions, latency, trajectories, and evaluation outputs.

A ROS 2 Humble mobile-robot workspace covering the main navigation software chain:

URDF / Xacro → Gazebo → ros2_control → Nav2 → waypoint navigation → autonomous patrol

It includes robot description, differential-drive control, localization/navigation configuration, Python navigation examples, custom services, patrol behaviors, and camera capture.

A deployment-oriented VLN scaffold connecting a Jetson-side ROS 2 robot with a remote GPU inference server. It covers camera history, remote NaVIDA inference, bounded action execution, /cmd_vel, and the STM32 serial chassis bridge.

• stronger VLN experimentation and evaluation methodology;
• LiDAR-aware navigation perception;
• Isaac Sim robotics workflows;
• deployment-aware embodied navigation on edge robotics platforms.