PhD Student
EPFL, Switzerland
Learning Algorithms and Systems
Laboratory (LASA)
Social
Researchgate
Contact:
Email: baiyu.peng@epfl.ch
Phone: +86 18810507500
Student projects I have supervised at the Learning Algorithms and Systems Laboratory (LASA), EPFL, exploring how robots can learn constraints from demonstrations and use them for motion planning and control.
Learns geometric constraint surfaces from demonstrated trajectories using a neural model. Projection-based corrections and optimal control then generate robot motions that stay on the learned surface while moving toward new goals.
Uses a large language model to propose relevant variables and constraint forms from a natural-language task, then fits numerical boundaries from a safe demonstration and synthetic unsafe examples. The resulting explicit inequalities are evaluated in a model predictive control trajectory planner.
Integrates constraints learned through Positive-Unlabeled Constraint Learning (PUCL) into a model predictive control planner as explicit inequalities. Simulation experiments study obstacle avoidance, generalization to new starts and goals, and how the planner's exploration affects constraint learning.
Investigates how to recover unknown state and action constraints from expert demonstrations when the task reward is known. The project alternates policy optimization and constraint inference, exploring neural classifiers and Gaussian mixture models to identify feasible regions.