Physical Human-Robot Interaction (pHRI)

The physical structure of a pHRI-capable robot is engineered to minimize injury risk during unexpected contact. This involves inherently safe kinematics (e.g., rounded edges, no pinch points) and back-drivable actuators that allow a human to easily move the robot. Materials are chosen to be compliant or padded. A key principle is low inertia, achieved through lightweight materials and motor placement, reducing the kinetic energy available in a collision. This design philosophy is foundational for collaborative robots (Cobots), distinguishing them from traditional industrial arms.

pHRI requires actuators that can sense and respond to external forces, moving away from rigid position control. This is achieved through:

• Series Elastic Actuators (SEAs): Motors connected via a physical spring, providing inherent force measurement and shock absorption.
• Torque-Controlled Motors: Direct control of joint torque, often using high-fidelity current sensors.
• Six-Axis Force/Torque (F/T) Sensors: Mounted at the wrist or base, these provide precise measurement of contact forces and moments applied to the end-effector. This suite of technologies enables impedance control and admittance control, allowing the robot to behave like a spring-damper system when touched.

The control system must detect unintended contact and react within milliseconds to prevent injury. This is implemented via:

• Model-Based Monitoring: Comparing expected joint torques (from the dynamic model) with measured torques. A discrepancy indicates external contact.
• Skin Sensors: Tactile arrays or capacitive sensors on the robot's surface provide direct contact location and pressure data.
• Predefined Safety Limits: Enforcing strict thresholds for speed, force, and power as defined in standards like ISO/TS 15066. Upon detection, reactions follow a hierarchy of safety: from a monitored stop, to retracting along the force vector, to triggering a protective stop that cuts power.

Beyond mechanical design, a robot's 'softness' or stiffness is primarily dictated by its control law. The two dominant paradigms are:

• Impedance Control: The robot regulates its dynamic relationship between position error and output force (Mass-Spring-Damper). The human feels a defined mechanical impedance when pushing the robot.
• Admittance Control: The robot regulates its motion in response to an applied force. A force sensor measures interaction, and the controller computes a resulting velocity or position change. These strategies enable physical guidance for kinesthetic teaching and allow the robot to 'yield' gracefully to human contact, which is essential for tasks like co-carrying an object.

For fluid collaboration, the robot must infer human intent directly from physical interaction, not just explicit commands. This involves interpreting:

• Guiding Forces: Differentiating between an intentional push to reposition the robot and an accidental bump.
• Haptic Cues: Recognizing patterns in force/torque signals that signify task phases (e.g., 'insert,' 'turn,' 'release').
• Ergonomic Load Sharing: Sensing how much weight or force the human is bearing during a co-manipulation task and adjusting its contribution accordingly. This moves interaction beyond simple 'hand guiding' into anticipatory assistance, where the robot begins to assist with a task based on the physical context.

pHRI systems must maintain a real-time model of the shared workspace to manage proximity and contact. This integrates:

• External Perception: Using depth cameras or LiDAR to track human position, posture, and velocity relative to the robot.
• Speed and Separation Monitoring (SSM): A safety function that dynamically adjusts the robot's speed based on the distance to the human, ensuring it can always stop before contact.
• 3D Safety Zones: Defining virtual volumes (using sensors or robot kinematics) that trigger specific behaviors (e.g., reduced speed in 'warning' zone, stop in 'protective' zone). This capability is critical for enabling proxemics-aware behavior and transitioning between different collaborative operation modes (e.g., hand guiding, power & force limiting).

Impedance Control regulates the dynamic relationship between a robot's position error and the output force, making the robot behave like a spring-damper system. Admittance Control inverts this relationship, mapping an applied force to a desired motion. These are foundational for physical collaboration, allowing a robot to yield to human push or provide compliant guidance.

• Key Mechanism: The controller implements a target mechanical impedance (stiffness, damping, inertia).
• Primary Use: Enabling kinesthetic teaching and safe response to unexpected contact.
• Example: A cobot sanding a part uses low impedance to follow surface contours, while high impedance is used for precise assembly.

Force/Torque Control directly commands the actuator forces or torques to achieve a desired interaction force with the environment or a human, rather than tracking a position trajectory. This is critical for tasks defined by force exchange.

• Key Mechanism: Uses force-torque sensors at the wrist or joint torque sensing for closed-loop force feedback.
• Primary Use: Assembly tasks (inserting a peg), polishing, or physical rehabilitation where consistent pressure is required.
• Challenge: Requires precise dynamic models and is less stable in free motion compared to position control.

Hybrid Position/Force Control decomposes a task into orthogonal subspaces: one controlled for position and another controlled for force. This allows a robot to simultaneously maintain a trajectory in some directions while regulating contact force in others.

• Key Mechanism: A selection matrix defines which degrees of freedom are under position or force control.

• Primary Use: Contact-rich tasks like wiping a surface (force normal to surface, position along it) or turning a crank.

• Foundation: Based on the task frame formulation, aligning control axes with the geometry of the task.

Series Elastic Actuation (SEA) is a hardware-level strategy where a compliant element (e.g., a spring) is intentionally placed in series between the motor and the robot link. This provides inherent force sensing, energy storage, and shock absorption.

• Key Mechanism: Motor controls spring deflection, which is measured to infer output force; the spring naturally filters high-frequency impacts.
• Primary Use: Enabling safe dynamic interaction in legged robots, prosthetics, and collaborative arms. It is a key enabler for force control without expensive joint torque sensors.
• Benefit: Fundamentally limits transient contact forces, a core principle of Power and Force Limiting (PFL) per ISO/TS 15066.

Variable Impedance Control dynamically adjusts the stiffness, damping, and inertia parameters of an impedance controller in real-time based on the task phase or sensory context. This enables a robot to switch between rigid precision and soft compliance.

• Key Mechanism: Parameters are modulated by a high-level task planner or learned policy.
• Primary Use: Assembly sequences (compliant search, then rigid insertion) or physical cooperation where the robot must adapt to unpredictable human motions.
• Advanced Form: Energy-aware control that minimizes metabolic cost for a human partner during co-manipulation.

Collision Detection and Reaction is a safety-critical software layer that identifies unintended contact and triggers a pre-defined safety response. It often works without dedicated skin sensors by using model-based observers.

• Key Mechanism: Compares expected joint torques (from a dynamic model) with measured currents/torques. A significant deviation indicates a collision.
• Reaction Strategies: Immediate torque cutoff, triggering a safety-rated monitored stop, or executing a reflexive retreat motion.
• Standard: This capability is a fundamental requirement for collaborative operation as defined in robot safety standards like ISO 10218 and ISO/TS 15066.

Shared Autonomy is a control architecture that dynamically blends human input with robot autonomy. Instead of binary manual/automatic modes, it creates a continuous spectrum of control allocation, crucial for intuitive pHRI.

• Mechanism: The system fuses human commands (from a joystick, gesture, or force sensor) with the robot's own planning algorithms to produce the final actuator commands.
• Key Techniques: Includes virtual fixtures (software guides), admittance control (motion in response to force), and intent recognition.
• Benefit: Allows the human to provide high-level guidance or make fine corrections while the robot handles low-level stability, precision, or obstacle avoidance.

Admittance Control and Impedance Control are two fundamental control strategies that define how a robot reacts to physical interaction, making it feel compliant and safe to touch.

• Impedance Control: The robot regulates the relationship between its position and the force it exerts (like a programmable spring-damper). It is good for stable, gentle contact.
• Admittance Control: The robot regulates its motion in response to measured forces (force in, motion out). It often feels more natural for hand-guiding but requires precise force sensing.
• pHRI Application: These strategies are what allow a cobot to 'give way' when pushed by a human or to apply a consistent, gentle force during an assembly task.

Learning from Demonstration (LfD), or Imitation Learning, is a primary method for programming robots in pHRI contexts. It allows non-expert users to teach tasks through natural demonstration rather than code.

• Kinesthetic Teaching: The user physically grabs and moves the robot arm through the desired task, recording joint angle trajectories.
• Teleoperation Demonstration: The user operates the robot via a controller to demonstrate the task.
• Outcome: The robot learns a policy (a mapping from state to action) to reproduce the skill. This is essential for deploying flexible cobots that can be quickly re-taught new tasks on the factory floor.

Intent Recognition is the perceptual challenge of inferring a human's immediate goals and planned actions from observable signals. In pHRI, this enables proactive and fluid collaboration.

• Input Signals: Uses force sensing (how the human is pushing/pulling), motion tracking (trajectory of the human's hand or tool), gaze tracking, and muscle activity (via EMG sensors).
• Algorithms: Often employs probabilistic models (e.g., Hidden Markov Models) or deep learning to predict the most likely intended action from a sequence of observations.
• Use Case: Allows a robot handing a tool to pre-position it based on the human's reaching trajectory, or to anticipate a human's need for a component during an assembly.