Model Predictive Control for Bipedal Robots

MODEL PREDICTIVE CONTROL ON A BIPEDAL ROBOT (seen to the right)

Bipedal walking robots are nonlinear dynamic systems that require close examination of the contact forces between the environment and the robot’s point of contact. However, past research has shown that it is difficult and limiting to rely solely on the contact forces to achieve stable postures. Yet, traditional tracking control systems (e.g., PD control) not only are structurally unable to deal with such dynamics, but they rely on predefined motions and adapt poorly to dynamic environments. This is especially problematic when the robot experiences perturbations either due to a force applied to the robot, the unpredictability of the ground, or noise from erroneous sensor data.

Thus, to achieve optimal bipedal walking, it is necessary to generate motions online and continually satisfy the dynamic state of the robot. Model Predictive Control (MPC) fulfills this need in the following ways: it provides a control strategy that computes optimal control values and corresponding state values online, it predicts future outputs over a predefined time horizon, and lastly, it minimizes its cost-function using specified constraints (i.e., this function may ensure that the Center of Mass or CoM and Center of Pressure are optimized for stability).