Event

RoKON: Adaptive Predictive Control Without Learning Explicit Models

1 reports · 1 sources

Akshun Sharma and other authors submitted a preprint proposing RoKON, a data-driven predictive control method for robotic manipulation that does not learn an explicit dynamics model. The method uses a Gaussian-kernel nonlinear predictor to represent the input-output data relationship, employs a damped Gauss-Newton method that exploits the least-squares structure of the control problem to achieve real-time control, and uses a hybrid buffer management strategy to incorporate informative online data while retaining relevant offline data, so as to adapt to changes in dynamics. The authors demonstrated tracking performance and computational suitability in MuJoCo using a Franka Research 3 robotic arm pushing blocks; the experiments did not know the changes in block friction values in advance. The paper states that these extensions pave the way for the method's use in real-time robotic tasks; the above results come from simulation demonstrations and have not been validated on a real robot.