Wenjun C.

Key Skills

Software

Don't disclose

Top Subject Matter

Robotics perception and multi-robot task execution (cloud-edge orchestration)

Multimodal robotic perception and instruction-driven robotic arm control

Imitation learning for embodied action prediction on robotic arm (teleoperation dataset)

Top Data Types

3D Sensor

Text

Image

Top Task Types

Fine-tuning

Segmentation

Classification

Freelancer Overview

Independent, Advisor: Prof. Yu Hao — Cloud-Edge-Device Collaborative Multi-Robot Agent Framework (Thesis). Brings 3+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Don't disclose, HuggingFace LeRobot, and PyTorch. Education includes Master of Science, Nanyang Technological University (2027) and Bachelor of Engineering, Southern University of Science and Technology (2026). AI-training focus includes data types such as 3D Sensor, Text, and Image and labeling workflows including Fine-tuning, Segmentation, and Classification.

Labeling Experience

Research Lead, Advisor: Prof. Yu Hao — WALL-X Based SO-101 Robotic Arm Imitation Learning System

3D SensorFine-tuning

Built a WALL-X based imitation learning system for a robotic arm that uses teleoperation-derived datasets to fine-tune a vision-language-action model. Assembled and debugged the SO-101 robotic arm hardware and integrated a leader-follower bilateral teleoperation setup with the LeRobot ecosystem to gather training data. Deployed WALL-X (WALL-OSS) using a Qwen2.5-VL backbone with a Flow Matching action prediction head and performed LoRA fine-tuning on self-collected teleoperation datasets. • Established a full imitation learning workflow from teleoperation data collection to hardware validation. • Deployed and fine-tuned the WALL-X model on self-collected datasets for precise object placement. • Validated the learned policy on real SO-101 hardware for end-to-end robotic tasks. • Achieved vision-language-action imitation learning using LeRobot and HuggingFace tooling.

2026 - 2026

Independent, Advisor: Prof. Yu Hao — Cloud-Edge-Device Collaborative Multi-Robot Agent Framework (Thesis)

Don't disclose3D SensorFine-tuning

Designed and deployed a cloud-edge-device multi-robot agent framework that relies on collecting and operationalizing perception and state data to support downstream decision-making and execution. Implemented unified heterogeneous communication adapters and periodic anomaly detection with checkpoint-resume recovery to maintain data and task continuity. Used depth-camera perception with 3D-to-2D dimensionality reduction to support navigation and robotic arm grasp execution for coordinated tasks. • Deployed Intel RealSense D455 depth sensing for calibration and state representation. • Built periodic state-polling and anomaly-driven checkpoint-resume logic under weak network. • Enabled Pure Pursuit autonomous navigation and arm grasp microservices driven by shared perception outputs. • Reported quantitative performance targets including vision ranging error and docking deviations.

2025 - 2026

Research Lead - Southern University of Science and Technology

ImageSegmentation

Developed an intelligent robotic arm control system using GLM-4V and SAM to improve physical coordinate accuracy. Fused semantic understanding and planning from a multimodal model with high-precision segmentation to improve zero-shot object recognition and grasping. Created a custom instruction set and optimized task decomposition via multi-agent prompt engineering with coordinate calibration and fault detection. • Built a perception-decision-execution closed loop mapping natural language input to precise grasping • Integrated GLM-4V semantic planning with SAM segmentation to reduce coordinate inaccuracy • Developed a custom high-level instruction set to expand robotic arm action decision space • Implemented coordinate calibration, dead-loop detection, and improved execution efficiency by prompt-driven decomposition

2024 - 2025

Core Developer (Summer School Project) - University of Oxford

ImageClassification

Built an ensemble image classification solution using transfer learning for a short summer school project. Developed a PyTorch preprocessing pipeline with data augmentation to improve generalization and model robustness. Trained and fine-tuned multiple CNN backbones by fusing multi-scale features into a customized embedding for evaluation. • Implemented PyTorch preprocessing with augmentation (random crop, flip, normalization) • Combined ResNet152, Inception v3, and Xception using transfer learning and multi-scale feature fusion • Modified classification heads and fine-tuned using a 6,144-dimensional feature representation • Tuned training with Adam optimizer and ReduceLROnPlateau scheduling to reach strong test accuracy

2024 - 2024

Core Developer, Oxford Summer School — AI & ML — Ensemble CNN-Based Image Classification with Transfer Learning

TextFine-tuning

Built and trained an ensemble CNN image classification model using transfer learning with a standardized PyTorch preprocessing pipeline. Applied data augmentation including random crop, flip, and normalization to improve generalization of the training dataset. Combined ResNet152, Inception v3, and Xception by freezing backbones, removing original classification heads, and fusing multi-scale features into a customized embedding for fine-tuning. • Implemented PyTorch preprocessing and augmentation for image inputs. • Performed transfer learning with backbone freezing and customized classification head via feature fusion. • Used Adam optimizer with ReduceLROnPlateau learning rate scheduling. • Achieved 91% test accuracy on the evaluation dataset.

2024 - 2024

Show more

Education

Work History