Whole-Body MPC vs Hierarchical WBC

Whole-Body Model Predictive Control (MPC) excels at generating dynamically optimal, coordinated motion because it treats the robot's entire state as a single optimization problem. By solving for contact forces and joint torques simultaneously over a receding time horizon, it produces fluid, reactive behaviors. For example, frameworks like Crocoddyl or Acados can compute trajectories that seamlessly blend a 1.5 m/s walking gait with a torso twist to reach for an object, achieving state-of-the-art energy efficiency in push recovery.

Hierarchical Whole-Body Control (WBC) takes a different approach by decomposing the problem into a strict task priority stack. High-criticality tasks like maintaining dynamic balance are solved in a null-space projection, ensuring they are never violated by lower-priority manipulation tasks. This results in deterministic safety guarantees. Implementations using OpenSoT or TSID allow a humanoid to lock its center of mass while performing a high-precision peg-in-hole insertion, even if the arm motion would otherwise destabilize a flat-footed stance.

The key trade-off: If your priority is dynamic agility and reactive planning in unstructured environments, choose Whole-Body MPC. If you prioritize modular safety decomposition and strict task guarantees for predictable industrial co-manipulation, choose Hierarchical WBC. The former leverages raw computational power for optimality, while the latter leverages software architecture for provable constraint enforcement.

Whole-Body MPC excels at generating highly dynamic, coordinated motions because it treats the entire robot as a single optimization problem. By solving for all joint torques and contact forces simultaneously, it can exploit the full-body dynamics to recover from aggressive pushes or execute fluid, athletic maneuvers. For example, implementations using Crocoddyl or Acados can achieve update rates of 200-500 Hz on torque-controlled humanoids, enabling reactive locomotion over rough terrain that hierarchical approaches struggle to match.

Hierarchical WBC takes a different approach by decomposing the problem into a strict priority stack—often placing the center of mass and swing foot tracking above postural tasks. This results in a modular, predictable, and computationally lighter framework that is easier to debug and tune in real time. The trade-off is a potential loss of optimality; when multiple constraints conflict, lower-priority tasks are simply sacrificed, which can lead to suboptimal whole-body coordination during highly dynamic maneuvers where every limb needs to work in concert.

The key trade-off: If your priority is pushing the limits of dynamic agility, energy efficiency, and robust push recovery in unstructured environments, choose Whole-Body MPC. If you prioritize deterministic, safety-certifiable behavior with predictable task decomposition and lower onboard compute requirements, choose Hierarchical WBC. Consider a hybrid approach where MPC generates optimal trajectories fed as reference inputs to a reactive WBC stack for industrial deployment.