Right Model, Right Time: Real-Time Cascaded-Fidelity MPC for Bipedal Walking

A team from the German Research Center for Artificial Intelligence (DFKI) has developed a real-time cascaded-fidelity model predictive control (MPC) framework for bipedal walking that intelligently balances precision and computational efficiency. The controller employs two models of differing fidelity: a detailed whole-body model for the near horizon (where accurate dynamics matter most) and a simplified single-rigid-body model for later prediction steps. This approach reduces the computational burden of solving a full nonlinear optimal control problem at every timestep, making it suitable for real-time deployment on hardware. The problem is solved using sequential quadratic programming (SQP) in the acados software package, and the controller plans joint torques directly without depending on pre-selected footstep locations—only requiring a target walking speed and a prior contact schedule.

Validation was performed in the MuJoCo physics simulator using the HyPer-2 bipedal robot, which has 18 degrees of freedom. The results demonstrate stable, efficient walking gaits that adhere to real-time constraints. The paper has been accepted to the IEEE ICRA 2026 Workshop on Frontiers of Optimization for Robotics. This work represents a meaningful step toward practical, computationally tractable whole-body control for humanoid robots, reducing the gap between simulation and real-world deployment.