Ph.D. student, Peking University
School of Computer Science · Open Lab on Human Robot Interaction
My research focuses on 3D perception, world-action models, and simulation for embodied AI, aiming to connect scene understanding with robotic manipulation. I study geometric and semantic representations for 6D object pose estimation, including generalization to unseen objects and the use of relationships between objects. I also build language-guided simulation pipelines for synthetic data generation and develop perception-guided dexterous grasping prototypes on real robots.
Open to research internships in world models and embodied AI.
Learning robot actions and future geometric states from sparse 3D robot–object skeletons. The model achieves 85.9% success on LIBERO-Plus with 57.1M parameters.
My contribution: simulation setup, training and tuning, robustness evaluation, ablations, analysis, writing, and technical figures. The metrics above are team paper results.
Connecting vision–language semantics with object geometry and spatial relationships for zero-shot 6D object pose estimation.
My contribution: led the research, method design, implementation, training, and evaluation.
Accepted papers (forthcoming)
Training-free pose estimation for unseen objects using RGB images and CAD-rendered templates. Template poses guide geometric matching, reaching 46.2% mean Average Recall across seven BOP datasets.
My contribution: contributed to the design of the proposed method, prepared figures illustrating its key ideas, and refined the manuscript’s technical content and presentation.
Estimating pose and size for new instances of known categories from RGB-D, using category-guided semantic and geometric features. Reaches 60.6% mAP at 5° / 2 cm on REAL275.
My contribution: contributed to defining the research motivation, carrying out ablation experiments, and refining the manuscript’s technical explanations and overall presentation.
Estimating multiple objects’ poses from a single RGB image and posed references, without CAD models or query depth. Reusable object memories and target-wise routing let difficult objects draw on reliable neighbors while retaining independent estimates as candidates.
Independently architected and implemented an LLM-assisted Blender–Unreal Engine pipeline for physics-validated scenes, RGB-D images, instance masks, calibrated cameras, and object 6D pose annotations.
Earlier controlled Phase 5B batch: 50/50 scenes, 1,000 annotated frames, six object categories. See the project page for the latest research progress.
Integrated RGB-D perception and scripted grasping on a Unitree G1-EDU with an Inspire RH56E2 hand.
Box placement: a reported ~47% reduction in routine time.
A research knowledge hub connecting geometric foundations, method taxonomies, datasets, evaluation metrics, and a searchable paper library for 6D object pose estimation.
Ph.D. in Computer Science and Technology
School of Computer Science
Intelligent Science and Technology
School of Artificial Intelligence and Data Science