Robot learning · Deformable object manipulation
✦ ICRA 2026 Transfer
Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control
Teaching robots to shape objects that are hard to see and predict.
Elasto-plastic objects such as clay can bend, stretch, and retain new forms. Severe self-occlusion and complex deformation dynamics make their state difficult to represent—and their motion difficult for a robot to plan.
The framework
3D occupancy
A volumetric state representation inferred from multiple RGB views captures the object beyond partial surface observations.
Learned dynamics
A model combining 3D convolutional and graph neural networks predicts complex object deformation.
Predictive control
A shape-aware action initialization module improves planning efficiency toward a desired goal shape.
Architecting a learning-based predictive control framework.
I contributed to the architecture of the predictive control framework and its 3D occupancy-based state representation during my research at CUHK.
Read the publication ↗Publication
Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control
Zhen Zhang, Xiangyu Chu, Yunxi Tang, Lulu Zhao, Jing Huang, Zhongliang Jiang, and K. W. Samuel Au
IEEE Robotics and Automation Letters, 2025 · Vol. 10, No. 7, pp. 7222–7229
ICRA 2026 Transfer
DOI: 10.1109/LRA.2025.3575308