Whole-body control (WBC) is the discipline of coordinating every joint in a humanoid robot as a unified system rather than controlling arms, legs, and torso independently. When a humanoid reaches for an object on a high shelf, it must simultaneously extend its arm, shift its center of mass, adjust its stance for balance, and possibly take a step — all in a fraction of a second. WBC frameworks manage these competing demands through optimization algorithms that resolve priorities in real time.
Traditional WBC approaches use hierarchical quadratic programming, where tasks are ranked by priority — balance first, then the manipulation objective, then posture preferences. Boston Dynamics' Atlas exemplifies this approach, seamlessly blending locomotion with dynamic manipulation. More recently, learning-based whole-body controllers trained via reinforcement learning in simulation have shown impressive results, with companies like 1X Technologies and Unitree demonstrating humanoids that walk, pick up objects, and recover from perturbations using end-to-end neural network controllers.
The transition from classical to learned WBC is one of the most active research frontiers in humanoid robotics. Hybrid approaches that combine the reliability of model-based control with the adaptability of learned policies are gaining traction, offering a practical path toward robots that can operate robustly in unpredictable real-world settings. For deeper coverage, see HumanoidIntel.