Accelerated gradient descent algorithm for MPC
Designed and Developed an accelerated gradient descent algorithm for real-time, high-frequency Nonlinear Model Predictive Control, achieving performance comparable to Differential Dynamic Programming (DDP) at 1 kHz. (LBR paper accpeted by UR 2024)
Can first-order methods achieve the same performance as (established) second-order methods when used at high frequency in MPC?
Key Words: Non-linear MPC, DDP, Accelerated Gradient Descent, C++, Real-robot deployment
- Enhanced solver efficiency by systematically investigating various first-order optimization algorithms for non-linear MPC problems, resulting in an algorithm that is 4x faster than DDP per iteration
- Reduced computation time to less than 1ms per control cycle for a 7-DoF manipulator by implementing advanced acceleration techniques (e.g. parallelization) in C++, enabling development on real-world manipulators at high frequencies
- Deployed the algorithm on a real-world manipulator, achieving performance comparable to DDP at 1kHz, while robust to external disturbances.