Whole-Body Control

Whole-body control is the coordination of all joints and actuators of a humanoid robot to achieve unified locomotion and manipulation. It is the foundational control layer that makes loco-manipulation possible — without whole-body control, arms and legs operate as independent subsystems, unable to leverage the full kinematic chain for tasks like reaching while walking or maintaining balance during heavy manipulation.

Gu et al.'s survey emphasizes the core trade-off in whole-body control: model fidelity versus computational efficiency. High-fidelity models (full rigid-body dynamics, contact mechanics, actuator dynamics) produce more accurate motions but are too slow for real-time control. Simplified models (inverted pendulum, centroidal dynamics) enable real-time performance but sacrifice accuracy in complex contact scenarios. Three decades of research have explored this trade-off without a clear winner — the optimal approach remains task-dependent.

Wen et al.'s Chain of Action framework includes a whole-body movement inference component that ensures kinematic and dynamic feasibility of generated behaviors. When the foundation model proposes a task decomposition, the whole-body inference layer checks whether the proposed motions are physically executable given the robot's joint limits, torque constraints, and balance requirements. This acts as a physics-aware filter between semantic task planning and motor execution.

Key Claims

• Model fidelity vs. computational efficiency is the core trade-off — Gu et al. document three decades of approaches without a clear resolution; high-fidelity models are too slow for real-time, simplified models sacrifice accuracy. Evidence: strong (Humanoid Locomotion & Manipulation Survey)
• CoA whole-body inference ensures kinematic/dynamic feasibility — Wen et al.'s framework includes a physics-aware layer that validates whether proposed motions are executable within the robot's physical constraints. Evidence: strong (Humanoid-COA: Chain of Action)
• Without whole-body control, arms and legs cannot coordinate — It is the prerequisite for loco-manipulation; independent limb control produces suboptimal or unstable behaviors. Evidence: strong (Humanoid Locomotion & Manipulation Survey)

Open Questions

• Can real-time whole-body optimization handle contact uncertainty in unstructured environments?
• How should tactile feedback be integrated into whole-body controllers for contact-rich tasks?
• What is the right level of model simplification for different task types?
• How do learned whole-body controllers compare to optimization-based approaches in terms of robustness and generalization?

Related Concepts

• Humanoid Loco-Manipulation — The application that requires whole-body control
• Imitation Learning — Learning whole-body behaviors from demonstration rather than engineering them

Backlinks

Pages that reference this concept:

• Humanoid Loco-Manipulation
• Imitation Learning