Let AI choose its embodiment: Co-design of learning and mechatronics for fast reflexive robotics (CoLearnMech)

Humans effortlessly perform fast and reflexive movements — such as flicking, grasping, or reacting to unexpected contact — while robotic systems often appear slow, cautious, and clumsy in comparison. Achieving such agility remains a fundamental challenge in robotics.

In recent years, learning-based control methods have shown impressive progress, enabling robots to tackle tasks that were previously considered difficult or infeasible, particularly in highly nonlinear and contact-rich settings. These advances naturally raise a fundamental question:

Are current limitations rooted in learning algorithms or in the physical embodiment of robots themselves?

Traditionally, robot design and control are treated as largely decoupled problems. Mechanical design decisions are typically made first, while learning or control algorithms are subsequently optimized for a fixed embodiment. While this separation simplifies engineering workflows, it may also impose fundamental limitations on achievable performance.

This project explores the hypothesis that jointly optimizing embodiment and learning-based control can lead to more capable and agile robotic systems. Rather than viewing learning and embodiment in isolation, we investigate their mutual dependence — how learning shapes the embodiment, and how the embodiment, in turn, shapes learning — as a promising direction toward fast and reflexive robotic behavior.

The project investigates a co-design framework in which learning-based control and physical embodiment are optimized jointly. Starting from a formalized design objective, an inner loop focuses on learning a high-performing control policy for a given embodiment using learning-based control methods. An outer loop then adapts the embodiment parameters based on the performance achieved by the learned controller. By iterating between these two processes, the framework aims to explore promising combinations of embodiment and control, explicitly accounting for their interdependence.

• Methods and experience in learning-based control, with a focus on reinforcement and imitation learning.
• Black-box optimization techniques for data-efficient exploration of complex design spaces, including bayesian optimization.
• Hands-on experience in designing and building robotic systems for learning-based control.