MSc student in Mechanical Engineering · EPFL
I am an MSc student in Mechanical Engineering at EPFL, where I work with Sylvain Calinon and Auke Ijspeert. I joined EPFL in 2024. My research focuses on robot learning, with interests in motion planning, manipulation, and locomotion.
Research perspective
My work across planning, manipulation, and locomotion has shaped how I think
about the gap between learning a skill and building a system that works.
Learning from data that
covers difficult scenarios.
Representing behavior for
learning and control.
Turning learned policies
into physical capabilities.
The question connecting these experiences
I am developing a benchmark to better understand the capabilities of embodied systems.
Build systems that combine learned skills with feedback and adaptation, and remain controllable and open to human supervision.
Motion planning
We present a hierarchical approach that integrates the spatial extraction capabilities of neural networks with the robust convergence of numerical optimization, enabling efficient and stable trajectory generation in diverse environments.
Manipulation
We propose Spline Policy, a structured policy representation that models robot behaviors as spline trajectories rather than discrete action chunks, supporting flexible temporal evaluation, parameter-space constraints, and uncertainty propagation.
Locomotion
We learn a single reinforcement-learning policy for walking, swimming, and automatic transitions between them using proprioceptive feedback. A biologically inspired prior guides learning, while a system-level sim-to-real alignment pipeline accounts for sensing, actuation, mechanical backlash, and fluid interactions. We validate the policy on a physical salamander robot on flat and uneven terrain and in water.
MSc in Mechanical Engineering
Advisors: Sylvain Calinon and Auke Ijspeert
Research on structured policies for manipulation and amphibious locomotion.
B.E. in Automation
Advisor: Fei Gao
Worked on learning-based motion planning, combining neural predictions with numerical optimization for trajectory generation.