Torque Control

Unlike position or velocity control, torque control sends a direct current command to the motor driver, which is proportional to the desired output torque. This bypasses the inner current loop of a traditional cascaded controller, providing the lowest-level access to actuator dynamics. It enables:

• Precise regulation of interaction forces with the environment.
• Intrinsic compliance, as the motor acts like a programmable force source.
• Direct implementation of high-level force-based control laws, such as impedance or admittance control.

A core advantage of torque control is the passive mechanical compliance it provides. By controlling output force rather than position, a joint can naturally yield upon unexpected contact, reducing the risk of damage to the robot or its surroundings. This is critical for:

• Human-robot collaboration where safe physical interaction is required.
• Operating in unstructured environments where collisions are unpredictable.
• Handling fragile objects without complex force sensing exoskeletons.

Torque control is the essential actuator-level primitive required to implement higher-level impedance control (regulating the dynamic relationship between position and force) and admittance control (regulating the dynamic relationship between force and position). In these schemes, a virtual model (e.g., a mass-spring-damper system) calculates the required torque command to achieve the desired interactive behavior, which is then executed by the torque-controlled actuator.

Accurate torque control in simulation requires detailed actuator models that capture non-linear dynamics. Key modeled elements include:

• Motor constants (Kt) relating current to torque.
• Electrical dynamics of the motor windings and driver.
• Friction models (Coulomb, viscous, Stribeck) within the gearbox and bearings.
• Torque saturation limits based on thermal and electrical constraints.
• Ripple and cogging torque effects from the motor's magnetic design.

The performance of a torque controller is directly tied to the accuracy of the system identification process for the physical actuator. This involves calibrating model parameters like:

• Motor torque constant and winding resistance.
• Gearbox efficiency and reduction ratio.
• Rotor and load inertia.
• Friction coefficients across the operating range. Discrepancies between the model and real hardware lead to torque tracking errors, which degrade the performance of high-level force-based controllers.

In Sim-to-Real Transfer Learning, training reinforcement learning policies that output torque commands is highly effective. It allows policies to learn complex, contact-rich behaviors with inherent compliance. The simulation must provide:

• A numerically stable physics engine that accurately resolves contact forces.
• Realistic actuator models with noise and latency.
• Domain randomization over model parameters (e.g., friction, inertia) to create robust policies that transfer to physical hardware despite modeling inaccuracies.

Torque control is essential for precision assembly tasks where controlling contact force is more critical than precise positioning. This is vital for inserting components, tightening fasteners to a specific torque, or mating parts with tight tolerances.

• Peg-in-Hole Insertion: The controller applies a gentle search force, allowing the peg to align and seat without jamming.
• Electronic Connector Mating: Prevents damage to fragile pins by limiting insertion force.
• Bolt Torquing: Directly commands the target torque for consistent, specification-compliant fastening.

In shared workspaces, torque control provides intrinsic safety by enabling compliant robot behavior. Instead of rigidly following a trajectory, the robot can yield to unexpected contact, such as a human touch.

• Physical Guidance (Hand Guiding): An operator can directly move the robot arm by applying force; the torque sensors detect this intent and the controller allows compliant motion for easy teaching.
• Collision Detection and Reaction: Abnormally high joint torque readings signal a collision, triggering an immediate stop or retreat motion.
• Cobots (Collaborative Robots): Most modern cobots use joint-level torque sensing as a foundational safety feature to meet ISO/TS 15066 standards.

These material removal processes require maintaining a constant normal force against a contoured surface. Position control alone fails as part geometry and tool wear vary.

• Adaptive Surface Following: Torque control adjusts the end-effector's position in real-time to maintain the desired contact force.
• Compliant Tool Holders: Often combined with passive compliance devices (like remote center compliance units) for high-frequency chatter absorption.
• Consistent Finish Quality: Ensures uniform material removal regardless of minor part misalignment or robot path inaccuracies.

In dynamic walking or running robots, torque control at the leg joints is critical for terrain adaptation and stability. It allows the robot to modulate leg stiffness and absorb impacts.

• Impedance Control Implementation: Often built on top of a torque-controlled actuator to regulate the dynamic relationship between foot position and ground reaction force.
• Impact Absorption: During foot strike, the controller commands lower torque to allow compliance, preventing high shock loads.
• Energy-Efficient Gaits: Enables spring-like energy storage and return in tendons or series elastic actuators, mimicking biological locomotion.

Torque control enables dexterous manipulation of delicate or irregularly shaped items where a simple binary grip is insufficient.

• Adaptive Grasp Force: The controller uses torque feedback (often at the finger joints or wrist) to apply just enough force to hold an object without crushing it (e.g., fruit, eggs, plastic components).
• Slip Detection: A sudden drop in measured torque at constant position can indicate object slippage, triggering an automatic increase in grip force.
• Haptic Exploration: By lightly dragging fingers over a surface, torque variations can be used to infer texture or identify edges.

Training reinforcement learning policies for force-sensitive tasks in simulation requires accurate actuator and contact dynamics modeling. Deploying these policies on real hardware depends critically on high-fidelity torque control loops.

• Policy Output as Torque Commands: Modern RL for robotics often outputs desired joint torques directly, requiring a low-latency, high-fidelity torque controller to execute the policy's intentions.
• Bridging the Reality Gap: Inaccuracies in simulated joint friction, motor dynamics, or contact models manifest as errors in predicted torque. Real-world torque control must compensate for these sim-to-real discrepancies.
• Domain Randomization Target: Parameters of the torque control loop itself (like PID gains, latency, noise) are often randomized during simulation training to create robust policies.

Impedance control is a higher-level strategy that regulates the dynamic relationship between a robot's end-effector motion and the contact forces it exerts. Instead of directly commanding torque, it makes the robot behave like a programmable mass-spring-damper system. This is often implemented using a low-level torque control loop.

• Key Idea: Controls dynamic behavior (stiffness, damping) rather than a specific force or position.
• Application: Tasks requiring gentle interaction, like assembly or physical human-robot interaction.

An actuator model is a mathematical representation of a physical motor's dynamics within a simulation. It is essential for training torque control policies that will transfer to real hardware. The model includes:

• Electrical dynamics: Relationship between input voltage/current and generated torque.
• Mechanical dynamics: Inertia, gearing, and output saturation limits.
• Non-linear effects: Static friction (stiction), Coulomb friction, and viscous damping.

Accurate modeling of these effects, especially backlash and torque ripple, is critical for closing the sim-to-real gap in force control.

Inverse dynamics is the computation of the joint torques required to achieve a desired motion (accelerations) for a robotic system, given its kinematic structure and mass distribution. It is a core component of model-based control.

• Contrast with Torque Control: Inverse dynamics calculates the theoretical torques needed for a planned trajectory. A torque controller is the low-level system that executes those torque commands on the actual hardware, compensating for unmodeled dynamics and disturbances.
• Use Case: Used in feedforward control loops to provide the bulk of the required torque, allowing the feedback controller (e.g., PID) to handle smaller corrections.

A PID (Proportional-Integral-Derivative) controller is a ubiquitous feedback mechanism. In the context of torque control, a current loop (a form of PID) is often the innermost control loop within a motor driver.

• Proportional (P): Responds to the present error between desired and measured current/torque.
• Integral (I): Eliminates steady-state error by accumulating past errors.
• Derivative (D): Predicts future error based on its rate of change, improving damping.

For torque control, the setpoint is a desired current (proportional to torque), and the process variable is the measured motor current.

Proprioception is a robot's sense of its own body's state. For torque control, the key proprioceptive sensors are:

• Joint Torque Sensors: Directly measure output torque at the joint, providing the highest-fidelity feedback for closed-loop torque control.
• Motor Current Sensors: Indirectly estimate torque (since torque is proportional to current), used when direct torque sensors are unavailable.
• Encoders: Provide joint position and velocity, which are necessary for model-based torque calculations and friction compensation.

Simulating the noise, bandwidth, and latency of these sensors is vital for training robust policies.

A friction model mathematically represents the resistive forces within an actuator or joint. Accurate friction compensation is arguably the most critical challenge for high-performance torque control. Key components include:

• Static Friction (Stiction): The force that must be overcome to initiate motion from rest.
• Coulomb Friction: A constant force opposing motion, independent of velocity.
• Viscous Friction: A force proportional to velocity.
• Stribeck Effect: The velocity-dependent friction at very low speeds.

In simulation, models like the LuGre model are used to capture these complex, non-linear dynamics for more realistic actuator behavior.