I am a B.S. candidate in Computer Science at Fudan University with an average grade of 90/100, and a former exchange student at The University of Texas at Austin (GPA: 4.0/4.0).
I conduct research at MagicLab, Fudan University, under the guidance of Wenchao Ding. My research focuses on execution feedback, policy adaptation, and failure recovery for long-horizon robotic manipulation, using world-action models as a primary platform.
From January to June 2026, I was a Test Development Engineer Intern at ByteDance (Douyin R&D, Live Streaming & Media Stream), where I worked on LLM-powered QA efficiency tools.
I am interested in robot learning, world models, and vision-language-action policies. My current questions concern how execution feedback and corrective data help a policy recover from failures, preserve required task progress, and generalize to unseen failure states. I study these questions through imitation learning, policy adaptation, and controlled closed-loop evaluation, including whether predicted futures provide useful information beyond observed execution history.
B.S. Candidate in Computer Science, Shanghai, China
Selected for the Top-Notch Talent Experimental Program in Computer Science and Technology (拔尖人才试验班).
Average grade: 90/100.
Studying execution feedback, policy adaptation, and failure recovery for robotic manipulation at MagicLab under Wenchao Ding. Previously contributed to risk-map research for autonomous driving, published in IEEE Robotics and Automation Letters (RA-L) in 2026.
Exchange Student in Computer Science, Austin, Texas, USA
GPA: 4.0/4.0 (13 credits); earned A grades in all four courses: machine learning, computer vision, operating systems, and probability/random processes.
Conducted multi-task active perception research at RobIn Lab under Junhong Xu, contributing to benchmark design and baseline evaluation workflows.
Advised by Wenchao Ding
Developing execution-feedback experiments with LingBot-VA on LIBERO-Long. I use state replay and controlled interventions to distinguish changes in model outputs, the availability of feasible recovery actions, and actual closed-loop task completion.
My current research asks how task-state feedback and recovery-data adaptation can help the same policy revise actions at subtask handoffs while preserving required task progress. I am investigating the value of feedback beyond ordinary recovery training, and whether predicted futures add useful information. Preliminary successes are limited to specific states and stages; recovery on unseen failure states remains under investigation.
Jie Jia, Yaofeng Su, Zeyu Bao, Yun Hong, Bingzhao Gao, Zhongxue Gan, Wenchao Ding
IEEE Robotics and Automation Letters (RA-L), 11(4): 4457–4464, 2026
Paper / arXiv
Contributed to the Transformer-based risk-prediction framework, model training, debugging, and result verification at MagicLab under Wenchao Ding.
Research Assistant, advised by Junhong Xu
Built multi-task partially observable simulation benchmarks with unknown object properties and spatial locations. Developed baseline training and evaluation workflows for VLA policies to study imitation learning, reinforcement learning, and IL pretraining followed by RL fine-tuning. Reviewed prior work on manipulation exploration and long-horizon POMDP benchmarks.
Research Assistant, advised by Prof. Xuanjing Huang
Constructed, evaluated, and optimized reflection datasets to improve model reasoning. Synthesized self-critic data based on Llama3.1-8b-instruct and generated diverse error-to-correct reflection trajectories. Applied tree-search and reverse-reasoning strategies; evaluated on GSM8K.
Test Development Engineer Intern, Douyin R&D, Live Streaming & Media Stream, Shanghai
Developed three AI-assisted QA tools for routine monitoring, test-case risk checking, and alert summarization. Refactored and expanded internal inspection and test cases, improving monitoring accuracy from 89% to 100% and coverage from 95% to 98% within the evaluated scope.
Built an end-to-end Vision-Language-Action system using diffusion policy. Integrated CLIP encoder and U-Net DDPM/DDIM architecture for high-fidelity action generation. Implemented 16-step action chunking with a full ManiSkill2 data-train-eval pipeline.
Combined OpenStreetMap with Gaode Maps API for route planning and location search. Implemented hybrid point selection, detailed route display, and multi-road-type support with speed constraints.
Developed a modular Python toolkit for neural style transfer with strong maintainability. Built batch processing and a configurable training pipeline for reproducible experiments. Provided an interactive web interface for in-browser stylization.