Whole-Body Control (WBC)

The defining feature of WBC is its strict hierarchical task prioritization. Tasks are assigned different priority levels, and the controller solves for joint motions that satisfy the highest-priority task first, then projects lower-priority tasks into the null space of higher ones. This ensures critical constraints like balance or joint limits are never violated to achieve a secondary goal like reaching.

• Example: A humanoid's primary task is maintaining Center of Mass (CoM) over its support polygon. A secondary task of reaching for an object is only executed with the remaining motion that does not disturb balance.

WBC uses a single, comprehensive dynamic model of the entire robot (e.g., a floating-base model for humanoids) within its optimization. This allows it to account for the dynamic coupling between all body segments. Unlike simpler methods that treat the arm and base separately, WBC understands that moving the arm generates reaction forces and torques that affect the base's stability, and plans motions accordingly.

This formulation is typically expressed as a Quadratic Program (QP) that minimizes control effort or acceleration while satisfying task constraints and physical limits.

Robots like humanoids or mobile manipulators are kinematically redundant, meaning they have more joints than strictly required to achieve a task's degrees of freedom. WBC explicitly exploits this redundancy. The null space of the primary task Jacobian represents all the possible joint motions that do not affect the primary task. WBC uses this null space to execute lower-priority tasks, optimize for posture, or avoid obstacles without interfering with the main objective.

WBC frameworks are built to natively handle a wide array of hard and soft constraints as part of the optimization problem. These are not afterthoughts but core inputs. Common constraints include:

• Physical Limits: Joint position, velocity, and torque limits.
• Contact Constraints: Forces must remain within friction cones, and unactuated contacts (feet on ground) cannot pull.
• Dynamic Consistency: Motions must be feasible according to the robot's equations of motion.
• Self-Collision Avoidance: Geometric constraints to prevent the robot from hitting itself.

WBC is fundamentally an online, receding-horizon optimization solved at control-loop frequencies (typically 100-1000 Hz). It does not pre-compute a full trajectory. Instead, at each control cycle, it takes the current state estimate and desired task references, solves the QP, and outputs optimal joint accelerations, torques, or velocities for the immediate next step. This makes it highly reactive to disturbances, sensor feedback, and changing task goals.

WBC is deeply rooted in Operational Space Control principles. It formulates tasks not in joint space, but in task space (or operational space)—the Cartesian space where the task is naturally defined (e.g., the position of an end-effector or the robot's CoM). The controller computes the forces/accelerations needed in this task space and then maps them back to joint torques using the dynamically consistent inertia matrix. This provides more intuitive and decoupled control of task behavior.

The most common mathematical formulation for WBC, Hierarchical Quadratic Programming (HQP) solves a stack of constrained optimization problems. Each level corresponds to a task (e.g., maintain balance, reach for an object). The solver finds joint commands that satisfy the highest-priority task first, then projects the solution into the null space of that task to solve for the next priority, minimizing interference.

• Primary Solver: Uses Null-Space Projections to ensure lower-priority tasks do not disrupt higher-priority ones.
• Core Constraint: Typically includes robot dynamics equations and joint torque/velocity limits.
• Real-Time Requirement: Must solve at control-loop frequencies (e.g., 1 kHz), requiring efficient QP solvers like qpOASES or OSQP.

For humanoid and legged robots, maintaining dynamic balance is the supreme-priority task in WBC. The controller uses a Centroidal Dynamics Model to relate joint torques to the acceleration of the robot's center of mass and its angular momentum.

• Primary Task: Regulate Center of Pressure (CoP) within the support polygon or manage Angular Momentum.
• Secondary Tasks: While balancing, the robot can use remaining control authority for arm manipulation, gaze control, or stepping.
• Example: Boston Dynamics' Atlas uses WBC to stay upright while performing parkour, where balance tasks dominate but arm swing tasks are executed in the null space.

WBC is essential for robots where a manipulator arm is mounted on a mobile base (e.g., a rover or a wheeled robot). It coordinates base motion and arm motion to achieve end-effector goals while respecting the system's physical constraints.

• Task Stack: 1) Avoid tipping over (stability), 2) Navigate to goal (base velocity), 3) Position end-effector (arm IK), 4) Minimize joint movement (comfort).
• Key Benefit: The mobile base can reposition itself to extend the workspace of the arm, solving problems a fixed-base manipulator cannot.
• Implementation: Used in warehouse robots like those from Fetch Robotics for shelf picking, where the base moves to align the arm with the bin.

Robots like humanoids have many more joints than required for a primary task (e.g., a 7-DoF arm needs only 6 DoF for end-effector pose). WBC uses this kinematic redundancy to optimize secondary criteria without affecting the primary task's performance.

• Common Optimization Criteria:
  • Joint Limit Avoidance: Keep joints away from mechanical stops.
  • Energy Minimization: Prefer configurations that require less torque.
  • Obstacle Avoidance: Maximize distance between the robot's links and environmental obstacles.
  • Human-Like Posture: Optimize for ergonomics or perceived safety in HRI.

WBC seamlessly integrates force control tasks with position control tasks. This is critical for compliant manipulation where the robot must exert specific forces (e.g., polishing, assembly) while maintaining its overall posture.

• Hybrid Force/Motion Control: A task can specify a desired force in one Cartesian direction and a desired position in another.
• Application - Peg-in-Hole: A WBC stack might prioritize: 1) Maintain robot balance, 2) Exert a downward insertion force, 3) Achieve a lateral alignment position.
• Implementation: Uses operational space formulations, where force tasks are defined in the same task-space framework as motion tasks.

In dynamic environments, WBC frameworks incorporate sensor feedback at the control level to react in real-time. This differs from slower re-planning loops.

• External Disturbance Rejection: Uses force/torque sensors in the feet or wrists to detect pushes. A high-priority "balance" task instantly computes joint corrections to regain stability.
• Unexpected Contact: If an arm makes unintended contact, a collision detection signal can trigger a high-priority "zero-force" task at the contact point, making the limb compliant.
• Reactive Grasping: Adjusts whole-body posture based on tactile sensor feedback during a grasp to prevent object slip or tipping.

Task and Motion Planning (TAMP) is an integrated approach that combines high-level symbolic reasoning (deciding what actions to perform) with low-level geometric motion planning (deciding how to move). It solves complex manipulation problems requiring long-horizon sequences. For a robot making coffee, TAMP would sequence actions like "grasp cup," "move to grinder," "place cup," and then solve the geometric paths for each.

• Hierarchical Integration: Symbolic planners output discrete action sequences, which geometric planners convert into continuous joint trajectories.
• Constraint Satisfaction: Must satisfy logical preconditions (e.g., gripper must be empty to grasp) and physical constraints (e.g., collision-free paths).
• Relation to WBC: WBC executes the low-level, continuous motion trajectories generated by the TAMP pipeline, often handling dynamic constraints and real-time disturbances that the planner may not model.

Model Predictive Control (MPC) is an advanced control strategy where an internal dynamic model predicts the system's future states over a finite time horizon. At each control cycle, it solves an optimization problem to find the optimal sequence of control inputs, applies the first step, and then re-plans. This enables handling of constraints (like joint limits or obstacle boundaries) and reactivity to disturbances.

• Receding Horizon: Continuously re-solves the optimization based on new sensor feedback.
• Core Mechanism: Minimizes a cost function (e.g., tracking error, energy use) subject to system dynamics and constraints.
• Relation to WBC: MPC is often the computational engine within a WBC framework. The WBC's quadratic program (QP) that coordinates multiple tasks can be formulated and solved as an MPC problem, especially for dynamic tasks like balancing while walking.

Inverse Kinematics (IK) is the computational process of calculating the joint angles or displacements required to achieve a desired position and orientation (pose) of a robot's end-effector. It solves the non-linear mapping from Cartesian task space to joint configuration space.

• Fundamental Problem: For a robotic arm, given a target (x, y, z, roll, pitch, yaw) for the gripper, IK computes the required shoulder, elbow, and wrist angles.
• Challenges: Often multiple or infinite solutions exist (kinematic redundancy); solutions may not exist due to reachability limits.
• Relation to WBC: In WBC, IK is generalized and solved dynamically. Instead of a single end-effector pose, WBC solves for joint velocities or accelerations that satisfy multiple simultaneous task constraints (e.g., one hand reaching, the other stabilizing, feet maintaining balance), all weighted by priority. It is differential inverse kinematics at the velocity/acceleration level.

Impedance Control and Admittance Control are two dual strategies for achieving compliant robot behavior during contact. They define how a robot reacts to external forces.

• Impedance Control: The controller modulates the dynamic relationship between position error and output force. It makes the robot behave like a mass-spring-damper system with programmable stiffness (K), damping (B), and inertia (M). The robot accepts position commands and regulates its interaction force.
• Admittance Control: The controller uses a force/torque sensor to measure external contact. This force measurement is fed into a desired admittance model (the inverse of impedance) to generate a motion correction. The robot accepts force commands and regulates its motion.
• Relation to WBC: These compliance strategies are often implemented as specific tasks within the WBC hierarchy. For example, a high-priority impedance control task can be defined for the end-effector to ensure safe human interaction, while a lower-priority task manages reaching. WBC resolves the potential conflicts between the motion generated by compliance and other objectives.

Force/Torque (F/T) Sensing is the measurement of the three-dimensional forces (Fx, Fy, Fz) and torques (Tx, Ty, Tz) applied at a robot's wrist or end-effector. This is typically achieved using a strain-gauge-based sensor mounted between the last joint and the gripper.

• Critical Data: Provides direct measurement of contact forces, essential for tasks requiring delicate interaction like assembly, polishing, or physical collaboration.
• Applications: Enables force-controlled operations such as peg-in-hole insertion, surface following, and hand-guiding for cobots.
• Relation to WBC: F/T sensor feedback is a primary input to WBC. It allows the controller to:
  • Implement admittance control tasks.
  • Estimate and compensate for external loads dynamically.
  • Maintain balance for humanoids by measuring the Center of Pressure (CoP) from foot sensors. Without accurate F/T data, many whole-body compliant behaviors are impossible.

Quadratic Programming (QP) is a mathematical optimization technique for minimizing a quadratic objective function subject to linear equality and inequality constraints. The standard form is: Minimize (\frac{1}{2} x^T Q x + c^T x) subject to (A_{eq} x = b_{eq}) and (A_{ineq} x \leq b_{ineq}).

• Computational Core: QP solvers are the workhorse of most modern WBC implementations.
• WBC Formulation: The control objective (e.g., minimize joint accelerations or tracking error) is the quadratic cost. Task constraints (e.g., end-effector acceleration = desired acceleration) become linear equalities. Physical limits (joint torque limits, friction cone constraints for contact) become linear inequalities.
• Hierarchy via Constraints: Strict task priorities are enforced by solving a sequence of QPs or using specialized solvers like Hierarchical Quadratic Programming (HQP) or Null-Space projections. The solution (x) is typically the vector of joint accelerations or torques.