System Calibration

Kinematic calibration corrects errors in the geometric model of a robot's manipulator or linkage system. It involves precisely measuring the actual Denavit-Hartenberg (DH) parameters—link lengths, joint offsets, and twist angles—which define the robot's forward kinematics. Inaccuracies arise from manufacturing tolerances, gear backlash, and structural deflection.

• Purpose: Minimizes end-effector positioning error by refining the mathematical model that maps joint angles to Cartesian space.
• Method: Uses external metrology tools (laser trackers, coordinate measuring machines) to collect precise 3D position data of the end-effector across many poses, then solves for optimal DH parameters.
• Impact: Critical for high-precision tasks like assembly and machining. Uncalibrated robots can have millimeter-level errors, while calibrated systems achieve sub-millimeter accuracy.

Dynamic parameter identification estimates the inertial and frictional properties of a robot's links and joints to create an accurate dynamic model. This model predicts the torques and forces required for a given motion, accounting for mass, center of mass, and inertia tensors.

• Purpose: Enables high-performance model-based control techniques like computed-torque control and Model Predictive Control (MPC), improving tracking accuracy and stability, especially at high speeds.
• Method: The robot executes a specially designed excitation trajectory while recording joint positions, velocities, accelerations, and motor currents/torques. A regression technique (e.g., least squares) solves for the dynamic parameters.
• Key Parameters: Mass, first moment of mass (center of mass), inertia tensor, and viscous/Coulomb friction coefficients for each link.

This process determines the internal optical characteristics (intrinsics) and 3D position/orientation (extrinsics) of cameras relative to the robot or world. It is foundational for all vision-based perception.

• Intrinsic Calibration: Solves for the camera's focal length, principal point, and lens distortion coefficients. It transforms pixel coordinates into normalized image coordinates.
• Extrinsic Calibration: Determines the rigid transformation (rotation and translation) from the camera frame to the robot's base or end-effector frame (hand-eye calibration).
• Process: Uses a calibration target (e.g., checkerboard, Charuco board) with known geometry. The robot or camera captures the target from multiple views to solve for parameters using algorithms like Zhang's method.
• Application: Essential for visual servoing, 3D reconstruction, and accurate sensor fusion.

IMU calibration characterizes the biases, scale factors, and non-orthogonalities of accelerometers and gyroscopes. These low-cost MEMS sensors have significant manufacturing variances that cause drift and error if uncorrected.

• Gyroscope Bias: The sensor's output when stationary should be zero. A constant bias causes unbounded angular drift.
• Accelerometer Bias & Scale: Errors cause incorrect gravity vector estimation, affecting tilt and acceleration measurements.
• Method: Typically involves a multi-position static calibration for the accelerometer (collecting data at multiple known orientations) and a tumble calibration for the gyroscope. Advanced methods also estimate cross-axis sensitivities.
• Importance: A prerequisite for accurate state estimation and sensor fusion in filters like the Kalman Filter, which are used for robot localization.

Force/torque (F/T) sensor calibration establishes the linear relationship between the sensor's raw voltage outputs and the actual forces and torques applied in the sensor's coordinate frame. These sensors are integral to compliant control and manipulation.

• Process: Involves applying a series of known, precise loads (masses) to the sensor in different directions and recording the output. A calibration matrix is then computed.
• Key Considerations: Must account for the sensor's mounting and the weight of any attached tool (end-effector), which creates a constant bias. This is often done via a "tare" operation.
• Use Case: Enables sensitive tasks like assembly insertion, object weighing, human-robot collaboration, and hybrid force-position control.

This calibration focuses on the non-ideal behaviors of motors and their drivetrains. It maps commanded control signals (e.g., current, position) to actual physical outputs, addressing non-linearities between simulation and hardware.

• Current/Torque Constant: Relates motor current to output torque, which can vary.
• Friction Mapping: Characterizes static (stiction) and dynamic (viscous) friction in gears and bearings, which are often poorly modeled in simulation.
• Backlash & Hysteresis: Quantifies the dead zone in geared transmissions where input motion does not cause output motion.
• Stiffness Identification: For robots with series elastic actuators or flexible joints, this measures the spring constant.
• Purpose: Creates a more accurate actuation model for simulation, reduces tracking error, and is critical for impedance control and safe physical interaction.

System Identification is the process of building or refining a mathematical model of a physical system's dynamics by observing its input-output behavior. It is a precursor to calibration, providing the parametric model that calibration then adjusts.

• Purpose: To create an accurate dynamic model (e.g., mass, inertia, friction coefficients) for use in simulation or control.
• Method: Involves exciting the system with known inputs (e.g., motor commands) and recording the outputs (e.g., joint angles, velocities) to fit model parameters.
• Relation to Calibration: System identification provides the what (the model parameters), while calibration is the how (the process of adjusting sensors and actuators to match those parameters).

The Reality Gap is the fundamental discrepancy between the simulated environment used for training and the target real-world operating conditions. Systematic errors from uncalibrated hardware are a primary contributor to this gap.

• Causes: Includes simplified physics, sensor noise models, actuator latency, and unmodeled mechanical backlash.
• Impact: Causes performance drop when a policy transfers from sim to real.
• Mitigation: System calibration directly reduces the systematic, deterministic portion of the reality gap related to hardware inaccuracies.

Domain Randomization is a sim-to-real technique that trains a policy by exposing it to a vast range of randomized simulation parameters, encouraging robustness to unseen real-world variations that calibration cannot fully address.

• Core Idea: Instead of perfecting one simulation, create many varied simulations.
• Randomized Parameters: Can include visual properties (textures, lighting), physical dynamics (mass, friction), and sensor characteristics (noise, bias).
• Synergy with Calibration: Calibration minimizes systematic bias; domain randomization builds robustness against residual stochastic variation and unmodeled effects.

Hardware-in-the-Loop Testing is a validation methodology where physical robot hardware (e.g., actuators, sensors, compute) is integrated with a real-time simulation of the environment and dynamics. It is a critical stage for calibration verification.

• Process: The physical robot's controllers receive commands from and send sensor data to a simulated world running in hard real-time.
• Purpose: Tests the integrated hardware-software system in a controlled, repeatable setting before full physical deployment.
• Calibration Role: HIL testing often reveals calibration errors (e.g., actuator response lag, sensor scaling) that are not apparent in pure software simulation.

Zero-Shot Transfer is the deployment of a policy trained entirely in simulation onto a physical robot without any fine-tuning or adaptation using real-world data. Its success is highly dependent on the quality of system calibration and simulation fidelity.

• Goal: Achieve functional performance immediately upon physical deployment.
• Prerequisites: Requires a high-fidelity simulation and a precisely calibrated robot to minimize the reality gap.
• Challenge: Any residual calibration error (e.g., a 2% error in gear ratio) can compound through dynamics, leading to task failure.

A Digital Twin is a high-fidelity, continuously updated virtual model of a physical system. For robotics, it serves as the ultimate calibrated simulation, synchronized with the real robot's state and used for monitoring, prediction, and offline optimization.

• Beyond Static Calibration: A digital twin is live, often updating its model parameters based on real-time sensor data.
• Applications: Predictive maintenance, what-if scenario testing, and offline policy re-training.
• Foundation: An accurate digital twin is built upon an initial, thorough system calibration and ongoing system identification.