Overview / Usage

Humans acquire motor skills by deliberate practice. Robots, on the other hand, must be meticulously programmed for every new task. This project leverages insights from human motor control in order to enable robotic systems to acquire and perfect badminton strokes through deliberate practice.

Methodology / Approach

Machine learning and optimal control blend together in an adaptive system that can extract useful knowledge from experience gathered through interaction with an external environment. In particular, motor programs are represented by compositions of movement primitives learned through iterative optimisation. Execution of motor programs on the physical system is facilitated by application of stabilising feedback controllers. The complete system crucially relies on learning and adaptation to accommodate for the variety of situations that arise in the game of badminton.

Technologies Used

• Optimisation of non-convex composite functions is carried out using PyTorch
• Model predictive control and trajectory optimisation are handled by CasADi
• Object tracking is delegated to Optitrack
• Computations are performed on the latest Intel Core i7 processors

Other links

• Robot Skill Learning