Persistent Excitation

A signal is persistently exciting of order (n) if its autocorrelation matrix is positive definite. Formally, for a discrete-time signal (u(k)), the condition is: [ R_{uu}(n) = \frac{1}{N} \sum_{k=1}^{N} \phi(k) \phi^T(k) > 0 ] where (\phi(k) = [u(k), u(k-1), ..., u(k-n+1)]^T) is the regressor vector. This ensures the signal contains at least (n) distinct frequency components, providing enough independent information to identify (n) parameters.

The signal must excite a sufficiently broad frequency spectrum that covers the natural frequencies and dynamic modes of the system being identified. Key aspects include:

• Bandwidth Coverage: The input's frequency content must span the system's operational bandwidth.
• Spectral Lines: A persistently exciting signal cannot be a single pure sinusoid; it requires multiple frequency components.
• Power Distribution: Energy should be distributed across frequencies to avoid ill-conditioned estimation problems. Common signals like pseudo-random binary sequences (PRBS) or chirp signals are designed for this property.

Excitation must be maintained over a sufficiently long time horizon to allow transient effects to settle and to average out measurement noise. This involves:

• Experiment Length: The data collection period must be long enough relative to the system's slowest time constants.
• Signal Persistence: The exciting properties must hold throughout the identification experiment, not just in brief bursts.
• Ergodicity: The time-averaged properties of the signal (used in calculation) should be representative of its ensemble properties.

Persistent excitation is a necessary condition for parameter identifiability in linear systems. Without it:

• The regression matrix becomes rank-deficient or ill-conditioned.
• The least-squares estimator fails, producing infinite or highly uncertain parameter estimates.
• Certain dynamic modes remain unobserved in the data, making their associated parameters impossible to estimate. This is directly related to the system's observability from the given inputs and outputs.

For robotic system identification, inputs are not arbitrary signals but planned excitation trajectories. Design principles include:

• Dynamically Rich Motions: Trajectories must move all joints through ranges that generate significant inertial, Coriolis, gravitational, and frictional forces.
• Optimal Experiment Design: Techniques like D-optimality are used to maximize the determinant of the information matrix, improving the conditioning of the estimation problem.
• Safety Constraints: Trajectories must respect joint limits, velocity bounds, and self-collision avoidance while still being exciting.

Failure to achieve persistent excitation leads to critical failures in the system identification pipeline:

• Uncertain Parameter Estimates: Large confidence intervals or covariance in the estimated parameters.
• Poor Model Generalization: The identified model may perform well on the training data but fail to predict novel motions.
• Degraded Sim-to-Real Transfer: An inaccurately identified simulation model widens the reality gap, causing policies trained in simulation to fail on physical hardware. This directly impacts the success of calibration and residual modeling efforts.

The process of constructing mathematical models of dynamic systems from measured input-output data. It involves:

• Experiment design to generate informative data.
• Model structure selection (e.g., linear, nonlinear).
• Parameter estimation to fit the model to data.
• Model validation to test predictive accuracy. Persistent excitation of the inputs is a prerequisite for reliable identification.

The algorithmic process of inferring the unknown constant values within a system's mathematical model from observed data. For a robot, this includes parameters like:

• Inertia tensors and link masses.
• Viscous and Coulomb friction coefficients.
• Motor torque constants. Methods include least-squares regression (for linear-in-parameters models) and gradient-based optimization. The accuracy of estimation is directly dependent on the input signal providing persistent excitation.

A deliberately designed sequence of control inputs or robot motions engineered to persistently excite all relevant dynamic modes. Key design principles include:

• Spectral richness: Covering a wide band of frequencies relevant to the system's dynamics.
• Amplitude variation: Ensuring signals are large enough to overcome noise and nonlinearities.
• Condition number optimization: Designing trajectories that yield a well-conditioned regression matrix for stable parameter estimation. These trajectories are critical for data collection in system identification pipelines.

A system-theoretic property measuring how well the system's internal states can be inferred from knowledge of its external outputs over time. A system is observable if, by observing the outputs, the initial state can be uniquely determined.

• Critical for state estimation (e.g., Kalman filters).
• Related to, but distinct from, persistent excitation. Persistent excitation concerns input signals for parameter identification, while observability concerns the relationship between states and outputs for state estimation.
• In robotics, ensuring observability is necessary for accurately estimating quantities like velocity from position sensors.

The degree to which a simulation model accurately replicates the behavior, dynamics, and outputs of the real-world system. High-fidelity models are essential for effective Sim-to-Real transfer. Fidelity is assessed through:

• Quantitative validation against real-world data.
• Task-specific benchmarking of trained policies.
• Analysis of residual errors (unmodeled dynamics). Accurate system identification, enabled by persistent excitation, is the primary method for physics parameter calibration to improve baseline model fidelity.

The performance discrepancy between a policy or model trained in simulation and its performance when deployed on the corresponding real-world physical system. This gap arises from:

• Simulation bias and unmodeled dynamics.
• Sensor and actuator noise not present in sim.
• Domain gaps in visual appearance or physics. Persistent excitation during system identification helps minimize the dynamics component of the reality gap by ensuring the simulation's physical parameters are accurately calibrated to the real robot.