🧠 Chapter 3: Control Systems & Locomotion

弘益人間 (홍익인간) · Benefit All Humanity

3.1 Control Architecture Overview

The control system of a humanoid robot consists of multiple hierarchical layers, each operating at different time scales and levels of abstraction. WIA-ROB-019 defines a standard control architecture that ensures safe, responsive, and intelligent robot behavior.

3.1.1 Hierarchical Control Layers

The standard recognizes four primary control layers:

Layer 1: Low-Level Joint Control (1-10 kHz)

The innermost control loop operates at the highest frequency, controlling individual joint motors. This layer implements PID controllers, torque controllers, or more advanced schemes like computed torque control. It runs on dedicated real-time processors with hard real-time guarantees.

Key responsibilities:

Layer 2: Whole-Body Control (100-1000 Hz)

This layer coordinates multiple joints to achieve desired body poses and trajectories. It implements balance controllers, center of mass regulation, and contact force distribution. Runs on real-time operating system with soft real-time constraints.

Key responsibilities:

Layer 3: Task Planning and Coordination (10-100 Hz)

Generates motion plans to accomplish high-level tasks. Includes path planning, obstacle avoidance, manipulation planning, and coordination between locomotion and manipulation. Can run on standard Linux with best-effort scheduling.

Key responsibilities:

Layer 4: Cognitive and Behavioral Control (1-10 Hz)

Highest-level decision making incorporating perception, natural language understanding, and goal-oriented planning. Integrates with WIA-INTENT for natural language commands and WIA-AI-CORE for machine learning models.

Key responsibilities:

Real-Time Guarantees

WIA-ROB-019 mandates that safety-critical control loops (Layers 1-2) run on real-time systems with deterministic scheduling. Layer 1 requires hard real-time (deadline misses never acceptable), while Layer 2 requires soft real-time (occasional deadline misses tolerable if infrequent). This ensures the robot maintains balance and safety even under high computational load from higher layers.

3.2 Balance and Stability Control

3.2.1 Center of Mass Control

Maintaining upright posture requires controlling the robot's center of mass (CoM) to remain within the support polygon defined by foot contacts. The fundamental balance condition is:

x_CoM ∈ Support Polygon

For static balance (no motion), the CoM projection must lie within the foot contact areas. During dynamic walking, we use the more sophisticated Zero Moment Point (ZMP) criterion.

Zero Moment Point (ZMP) Method

The ZMP is the point on the ground where the net moment of all forces acting on the robot equals zero. For stable walking, the ZMP must remain inside the support polygon. WIA-ROB-019 requires ZMP-based balance control for all Level 2 and higher compliance.

The ZMP position is calculated from the robot's dynamics:

ZMP = (Σ m_i × [x_i × ẍ_i + g]) / (Σ m_i × [ẍ_i_z + g])

Where m_i are segment masses, x_i are segment positions, ẍ_i are accelerations, and g is gravity.

Control strategies:

3.2.2 Fall Recovery Strategies

Despite sophisticated balance control, external disturbances may exceed the robot's balance capability. WIA-ROB-019 defines a hierarchy of recovery strategies:

  1. Ankle Strategy (small disturbances): Adjust ankle torque to shift CoM back over support. Effective for perturbations <5% body weight. Response time: <100ms.
  2. Hip Strategy (medium disturbances): Rapid hip flexion/extension to move CoM. Handles larger perturbations while maintaining same foothold. Response time: <200ms.
  3. Step Strategy (large disturbances): Take corrective step to establish new support polygon. Requires rapid step planning and execution. Response time: <500ms.
  4. Controlled Fall (unavoidable): If balance unrecoverable, initiate controlled falling sequence to minimize impact and protect critical components. Deploy arms to cushion fall, tuck head, and cut power to actuators.

🛡️ Safety Protocol

All humanoid robots must implement fall detection and controlled falling procedures. IMU signals are continuously monitored for tilt angles exceeding thresholds (typically 30-45°). Upon fall detection, the robot transitions to a protective mode: extending arms, reducing joint stiffness to absorb impact, and logging detailed sensor data for post-incident analysis.

3.3 Walking Pattern Generation

3.3.1 Gait Phases and Timing

Human walking consists of distinct phases that WIA-ROB-019 formalizes for robot locomotion:

Phase Duration (% of cycle) Key Events Control Objectives
Heel Strike 0% Leading foot contacts ground Impact absorption, prevent slipping
Double Support 0-10% Both feet on ground Weight transfer, ZMP transition
Single Support 10-50% One foot supports full weight Maintain balance, swing leg forward
Swing Phase 50-100% Leg moves forward Toe clearance, trajectory tracking
Opposite Heel Strike 100% (= 0% next cycle) Other foot contacts Cycle completion, repeat

Typical step cycle duration: 0.8-1.2 seconds at normal walking speed (0.8-1.2 m/s). Double support phase critical for stability—minimum 10% of cycle recommended.

3.3.2 Trajectory Generation Methods

Parametric Gait Synthesis

Uses analytical functions to generate joint trajectories based on gait parameters (step length, height, timing). Fast computation enables real-time adaptation. Common approaches:

Optimization-Based Planning

Formulate walking as an optimization problem minimizing energy consumption, torque, or other cost functions subject to dynamics and constraints. More computationally expensive but generates optimal motions for complex scenarios.

Typical formulation:

minimize   J = ∫ [w₁||τ||² + w₂||CoM - CoM_ref||²] dt
subject to  robot dynamics
            ZMP ∈ support polygon
            joint limits
            torque limits
            collision avoidance
        

Solution methods: Sequential Quadratic Programming (SQP), Interior Point methods, Direct Collocation. Computation time: 0.1-10 seconds depending on complexity and horizon length.

Learning-Based Approaches

Deep reinforcement learning and imitation learning show promise for generating natural, robust walking patterns. Trained in simulation then transferred to hardware. Benefits include adaptation to terrain and disturbances. Challenges include sample efficiency, safety during training, and sim-to-real transfer.

WIA-ROB-019 permits learning-based controllers but requires:

3.4 Terrain Adaptation and Compliance

3.4.1 Surface Detection and Classification

Robust locomotion requires adapting to surface properties. WIA-ROB-019 requires the robot to detect and classify terrain types:

Terrain Type Friction Coefficient Stiffness Required Adaptations
Concrete/Asphalt 0.7-0.9 Very High Standard gait parameters
Carpet 0.5-0.7 Medium Increased foot lift, moderate compliance
Tile (wet) 0.2-0.4 Very High Reduced speed, smaller steps, careful weight transfer
Grass 0.4-0.6 Low-Medium Higher foot lift, increased compliance
Gravel 0.6-0.8 Low Wide stance, slower speed, high compliance
Stairs (up) 0.7-0.9 Very High Specialized stair climbing gait
Stairs (down) 0.7-0.9 Very High Controlled descent, visual verification of steps

Detection methods:

3.4.2 Slope and Uneven Terrain Handling

Walking on slopes requires adjusting the robot's orientation and modifying gait parameters:

3.5 Manipulation and Whole-Body Coordination

3.5.1 Reaching and Grasping

Humanoid robots must coordinate arm motion with body posture and balance. WIA-ROB-019 specifies workspace requirements and coordination strategies.

Workspace Definition

The reachable workspace depends on arm kinematics and balance constraints:

Grasp Planning Pipeline

  1. Object Segmentation: Identify target object from visual data. Deep learning-based instance segmentation. Processing time: 50-200ms.
  2. Pose Estimation: Determine 6-DOF object pose. Methods include template matching, point cloud registration, learned pose estimators. Accuracy: ±5mm position, ±5° orientation.
  3. Grasp Synthesis: Generate candidate grasps considering object geometry, task constraints, collision avoidance. Analytical methods or learning-based (GraspNet, Contact-GraspNet). Computation: 100-500ms for 50-100 candidates.
  4. Grasp Ranking: Evaluate candidates using quality metrics (force closure, stability, manipulability). Select best feasible grasp.
  5. Motion Planning: Plan collision-free trajectory from current configuration to grasp pose. RRT, PRM, or optimization-based planners. Planning time: 0.5-5 seconds.
  6. Execution: Track planned trajectory while maintaining balance. Visual servoing for final approach to handle uncertainties.

3.5.2 Whole-Body Motion Planning

Many tasks require coordinated motion of legs, torso, and arms. Examples include:

WIA-ROB-019 recommends task-space optimization formulations that treat the entire robot as a kinematic chain:

minimize   ||q̇||² + w||ė - ė_desired||²
subject to  J(q)q̇ = ė_desired  (task constraint)
            ZMP ∈ support polygon  (balance)
            q_min ≤ q ≤ q_max  (joint limits)
            ||q̇|| ≤ q̇_max  (velocity limits)
        

Where J(q) is the task Jacobian, ė_desired is the desired end-effector velocity, and q̇ are joint velocities. Solved at each time step (100-200 Hz) using quadratic programming.

3.6 Compliance and Impedance Control

3.6.1 Impedance Control Framework

For safe physical interaction with humans and the environment, robots must exhibit compliant behavior. Impedance control shapes the relationship between applied forces and resulting motion:

F = M(ẍ - ẍ_d) + B(ẋ - ẋ_d) + K(x - x_d)

Where F is the interaction force, and M, B, K are desired inertia, damping, and stiffness matrices. By adjusting these parameters, the robot can be made more compliant (soft, back-drivable) or stiff (precise position control).

WIA-ROB-019 requires variable impedance control for safe human interaction:

3.6.2 Force/Torque Control

Direct force control enables tasks requiring specific contact forces:

Implementation requires accurate force sensing (F/T sensors) and force control loops running at >200 Hz. Control bandwidth typically 10-50 Hz depending on mechanical compliance.

3.7 Chapter Summary

This chapter explored the control systems that enable humanoid robot locomotion and manipulation. We examined hierarchical control architectures, balance control using ZMP methods, walking pattern generation, terrain adaptation, whole-body motion planning, and compliant interaction through impedance control.

These control strategies transform the mechanical hardware described in Chapter 2 into a functional system capable of stable walking, adaptive behavior, and safe human interaction. The WIA-ROB-019 control specifications ensure consistent performance across different robot platforms.

In Chapter 4, we'll investigate the sensor systems and perception algorithms that provide the data these controllers rely on, enabling robots to understand and navigate complex real-world environments.