Jinjia Guo

I am currently pursuing my Master degree of Robotics at the University of Michigan in Ann Arbor. There, I worked with ProfessorBernadette Bucher.

Previously, I earned dual Bachelor's degrees in Mechanical Engineering from the University of Cincinnati (UC) and Chongqing University (CQU). During my undergraduate studies, I also completed a minor in Robotics at UC. Throughout this period, I engaged in research under the mentorship of Professor Janet Jiaxiang Dong and Professor Rui Li. In addition, in 2022, I collaborated with Nan Mu and Professor Jingfeng Jiang at Michigan Technological University (MTU).

Email  /  CV  /  Scholar  /  Github

Research Interests

I am broadly interested in embodied intelligence at the intersection of robot learning, long-horizon planning, and contact-rich manipulation. Motivated by Moravec's paradox—that tasks trivial for humans can be surprisingly hard for machines—my long-term goal is to build robotic systems that can autonomously accomplish everyday tasks in realistic, unstructured environments.

Projects

We are building a photorealistic, physics-based benchmark for interactive object search, where a mobile manipulator must navigate multi-room homes, open articulated furniture, and grasp target objects based on high-level language instructions. The benchmark is constructed by procedurally generating Infinigen indoor scenes and importing them into IsaacSim, with articulated doors, cabinets, and drawers, and 200+ object categories.

We achieve robust in-hand screwdriver manipulation by tightly coupling vision-based point cloud perception with tactile force closure optimization, using diffusion models for trajectory proposals and online LP-based adjustments to maintain stable contact during rotation.

We propose a hybrid dynamics modeling approach, SAIN, that combines a physics engine with residual learning. By integrating it with MPPI control, we achieve 100% success rate in manipulating target objects under 40 diverse and randomized physical conditions.

This study compares A* and ANA* algorithms on the PR2 robot, showing that heuristic design critically balances planning efficiency and safety, with ANA* offering greater adaptability in dynamic environments while A* excels in real-time responsiveness.

In the Botlab, movement control, obstacle detection, maze exploration, and self-localization functionality was developed on the MBot robot, a mobile robot platform. It is designed to explore the fundamentals of robot autonomy by developing MBot with autonomous mapping, localization, and exploration capabilities.

In the Armlab, a 5-DOF robotic arm fully autonomously arranges blocks of different sizes, colors and positions into the desired arrangement. Numerical inverse kinematics is used to determine the appropriate waypoints. An overhead LiDAR Camera is utilized to identify blocks on the board.

In the Armlab, a 5-DOF robotic arm fully autonomously arranges blocks of different sizes, colors and positions into the desired arrangement. Numerical inverse kinematics is used to determine the appropriate waypoints. An overhead LiDAR Camera is utilized to identify blocks on the board.

Publications

We proposed MBNet, a novel network combining high-low feature aggregation and hierarchical supervision for salient object detection in nighttime scenes, achieving superior performance over seven state-of-the-art methods on a newly built low-light dataset.

Experiences

Applied RAG in JittorLLM, a large language code completion model based on the self-developed framework: Jittor. The model reduces latency by 70% and improves accuracy by 20% compared with Copilot.

Demonstrated strong written and verbal communication skills by providing clear, constructive feedback on assign- ments for a cohort of 103 students, ensuring high standards of professionalism and efficiency.

Designed and implemented a new salient object detection network tailored for medical imaging, utilizing multi-scale fusion in neural networks to achieve a performance improvement of over 10%.

Created a dataset for raw-material classification and segmentation and developed a Unet++ algorithm for granularity detection on belt machines for Magang (Group) Holding Co., Ltd.