Definition

Whole-body control (WBC) is the problem of coordinating all degrees of freedom of a robot simultaneously to achieve task-space objectives while satisfying physical constraints — balance, joint limits, contact forces, collision avoidance, and actuator torque bounds. Unlike arm-only control where the base is fixed, WBC must manage the coupling between locomotion and manipulation, keeping the robot balanced while its arms perform useful work.

For a humanoid robot with 30+ actuated joints, two legs making intermittent contact with the ground, and two arms reaching for objects, the WBC problem is formidable: the controller must decide, at every timestep, how to distribute forces across all contact points, which joints to move and how fast, and how to trade off competing objectives (e.g., tracking accuracy vs. balance margin) — all within a 1-5ms computation budget for real-time control.

Why It Is Hard

High dimensionality: A humanoid has 30-50+ DOF. A mobile manipulator (arm on a wheeled or legged base) has 10-20+ DOF. The optimization must simultaneously determine commands for all joints, producing a high-dimensional decision space that must be solved in real time.

Balance constraints: For legged robots, the center of mass (CoM) must remain within the support polygon defined by the feet in contact with the ground. Any manipulation motion that shifts the CoM outside this polygon causes the robot to fall. Balance is a hard constraint that cannot be violated, even momentarily.

Contact forces: Ground reaction forces at the feet must satisfy friction cone constraints (no sliding) and unilateral constraints (the ground can only push, not pull). These constraints are nonlinear and change as the robot's contact state changes (e.g., lifting a foot during walking).

Real-time requirements: WBC must run at 200-1000 Hz for stable control of dynamic systems. This leaves only 1-5ms per control cycle, demanding extremely efficient optimization solvers.

Underactuation: Legged robots are inherently underactuated — the floating base (the torso) has 6 DOF (position + orientation) but no direct actuation. The base can only be controlled indirectly through contact forces at the feet, making the control problem fundamentally different from a fixed-base manipulator.

Key Algorithms

  • Hierarchical Quadratic Programming (HQP) — The most widely used WBC formulation. Tasks are organized in a strict priority hierarchy: highest-priority tasks (balance, contact constraints) are solved first, and lower-priority tasks (end-effector tracking, posture regulation) are solved in the null space of higher-priority tasks. Each priority level is a QP, and the cascade ensures that critical constraints are never violated. Popularized by the Stack of Tasks (SoT) framework from LAAS-CNRS.
  • Weighted QP (soft priorities) — An alternative that combines all tasks into a single QP with weighted cost terms. Faster to solve (one QP instead of a cascade) but does not guarantee strict priority ordering — a heavily weighted balance task might still be slightly violated if the manipulation task has a high weight. Used when strict guarantees are less critical or when the priority structure is unclear.
  • Centroidal dynamics — A simplified dynamics model that describes the motion of the robot's center of mass and angular momentum as functions of contact forces, without modeling individual joint dynamics. Used for trajectory optimization and long-horizon planning, producing reference CoM and momentum trajectories that the WBC tracks at the joint level.
  • Model Predictive Control (MPC) — Optimizes a trajectory over a finite time horizon (0.5-2 seconds) while predicting future dynamics. MPC naturally handles dynamic motions (walking, jumping, throwing) where the robot's contact state changes over time. Convex MPC formulations (linearized dynamics) solve in 1-5ms; nonlinear MPC provides higher fidelity but requires more compute.
  • Operational Space Control — Khatib's framework for controlling end-effector motion while using the null space for secondary objectives (balance, posture, obstacle avoidance). The original formulation of multi-priority task-space control, extended by modern WBC methods.

Applications

Humanoid locomotion + manipulation: The primary motivation for WBC. A humanoid picking up a box from a shelf must coordinate arm reaching with body leaning, weight shifting, and stepping — all while maintaining balance. WBC enables this simultaneous coordination, which sequential (first walk, then reach) approaches cannot achieve efficiently.

Legged robot manipulation: Quadruped robots (Unitree Go2, Boston Dynamics Spot) equipped with arms use WBC to manipulate objects while maintaining four-legged stability. The legs adjust stance to counteract manipulation forces, and the controller manages the tradeoff between base stability and arm dexterity.

Mobile manipulators: Wheeled or tracked platforms with mounted arms (Fetch, TIAGo, Stretch) use WBC to coordinate base movement with arm motion, enabling longer reaches and more dexterous workspace coverage than fixed-base arms.

Dynamic motions: Acrobatic behaviors — jumping, running, backflips, throwing — require whole-body coordination that accounts for flight phases, impact dynamics, and momentum management. Boston Dynamics' Atlas and Unitree's H1 demonstrate these capabilities powered by WBC + RL combinations.

WBC and Learning-Based Control

Modern legged robot control increasingly combines learned policies with WBC in a hierarchical architecture:

RL policy as high-level planner: A neural network policy (trained in simulation via RL) outputs desired foot positions, base velocity, or CoM trajectory. A classical WBC converts these high-level commands into joint torques while enforcing physical constraints. This hybrid approach leverages RL's ability to learn complex behaviors while relying on WBC for physical feasibility and safety.

End-to-end learned WBC: Some approaches train neural networks to directly output joint torques, bypassing classical WBC entirely. This requires the policy to implicitly learn balance, contact force management, and constraint satisfaction. It works well in simulation and for locomotion (where the RL reward naturally penalizes falling), but remains challenging for manipulation where constraint violations can cause hardware damage.

Sim-to-real for WBC: Training WBC policies in simulation and transferring to real hardware is the dominant paradigm for legged locomotion. Domain randomization of dynamics parameters (friction, mass, motor characteristics) produces robust policies that transfer without fine-tuning.

Leading Research Groups

  • ETH Zurich RSL (Robotic Systems Lab) — Pioneered learning-based locomotion for ANYmal quadruped. Developed perceptive locomotion policies that traverse challenging terrain, combined with WBC for robust execution.
  • MIT Biomimetic Robotics Lab (Cheetah Lab) — Developed the MIT Cheetah and Mini Cheetah platforms. Pioneered convex MPC for dynamic locomotion at 30+ Hz control rates, enabling running, jumping, and backflips.
  • Boston Dynamics — Industry leader in dynamic whole-body control. Atlas demonstrates the most advanced humanoid locomotion and manipulation, using proprietary WBC + optimization-based planning.
  • LAAS-CNRS (Toulouse) — Developed the Stack of Tasks framework for hierarchical WBC, one of the most influential open-source WBC implementations. Applied to HRP humanoid platforms.
  • Unitree Robotics — Produces the G1 humanoid and Go2 quadruped. Combines RL-trained locomotion policies with WBC for increasingly capable whole-body behaviors at consumer price points.

Key Papers

  • Khatib, O. (1987). "A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space Formulation." IEEE Journal of Robotics and Automation. The foundational operational space control framework.
  • Sentis, L. & Khatib, O. (2005). "Synthesis of Whole-Body Behaviors through Hierarchical Control of Behavioral Primitives." IJHR. Introduces hierarchical whole-body control with prioritized task execution.
  • Di Carlo, J. et al. (2018). "Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control." IROS 2018. Demonstrates real-time convex MPC for dynamic quadruped locomotion.
  • Lee, J. et al. (2020). "Learning Quadrupedal Locomotion over Challenging Terrain." Science Robotics. ETH RSL's breakthrough in learning-based locomotion with sim-to-real transfer for ANYmal.

Related Terms

  • Inverse Kinematics — WBC generalizes IK to include dynamics, contacts, and constraints
  • Joint Space — The high-dimensional space WBC optimizes over
  • Safety Constraints — Balance and force limits are the primary safety constraints in WBC
  • Motion Planning — Provides reference trajectories that WBC tracks in real time
  • Sim-to-Real Transfer — The dominant paradigm for training WBC policies for legged robots

Apply This at SVRC

Silicon Valley Robotics Center operates humanoid and quadruped platforms (Unitree G1, H1, Go2) with whole-body control capabilities. Our team supports WBC development, RL policy training for locomotion and loco-manipulation, and sim-to-real transfer testing in our controlled lab environment. Whether you are developing custom WBC for a new platform or fine-tuning locomotion policies, we provide the hardware, compute, and expertise.

Robot Leasing   Contact Us