Dynamic Regressor

The defining property of a dynamic regressor is its linear-in-the-parameters structure. It expresses the complex, nonlinear equations of motion as a linear combination of unknown dynamic parameters. The standard form is:

Y(q, q̇, q̈) * Φ = τ

Where:

• Y(q, q̇, q̈) is the regressor matrix, a function of measurable joint positions (q), velocities (q̇), and accelerations (q̈).
• Φ is the vector of unknown dynamic parameters (e.g., link masses, inertias, friction coefficients).
• τ is the vector of measured or commanded joint torques.

This linearity transforms a nonlinear estimation problem into a linear least-squares problem, which is computationally efficient and globally convergent.

The regressor matrix Y is not arbitrary; it is systematically derived from the Newton-Euler or Lagrangian formulations of rigid-body dynamics. The process involves:

• Writing the full nonlinear dynamics equations for the multi-body system.
• Factoring out all terms that multiply each unknown base parameter.
• Structuring these terms into the columns of the Y matrix.

Key Insight: This derivation ensures the regressor encodes the fundamental physics of the system. The unknown parameter vector Φ typically contains minimum inertial parameters (also called base parameters), which are the unique combinations of physical masses and inertias that actually affect the observed dynamics.

The dynamic regressor is the engine of data-driven parameter estimation. The standard identification procedure is:

1. Execute a rich excitation trajectory on the real robot and record [q, q̇, q̈, τ] data.
2. Construct the regressor matrix Y for each timestep using the recorded q, q̇, q̈.
3. Stack data over time to form the linear system: [Y_stack] * Φ = [τ_stack].
4. Solve for Φ using linear least-squares (e.g., via the pseudo-inverse).

The accuracy of the estimated parameters Φ̂ depends critically on the persistent excitation of the collected data, which ensures the matrix Y_stack is well-conditioned for inversion.

Accurate parameter estimates Φ̂ are directly used to calibrate physics simulators, closing the reality gap. The process is:

• The estimated Φ̂ (masses, inertias, frictions) from the real robot are injected into the simulator's dynamic model.
• This parameter calibration minimizes the simulation bias in the forward and inverse dynamics predictions.
• A high-fidelity, calibrated simulator enables reliable Sim-to-Real Transfer for training reinforcement learning policies or testing controllers.

This makes the dynamic regressor a foundational component in the System ID Pipeline, bridging measured reality and virtual simulation.

The properties of the regressor matrix Y are deeply connected to fundamental system concepts:

• Persistent Excitation & Observability: For parameter estimation, the excitation trajectory must make the stacked Y_stack matrix full column rank. This is the persistent excitation condition, which is analogous to the observability condition for state estimation. It ensures all dynamic parameters are "visible" in the collected data.
• Controllability for Excitation: Designing an effective excitation trajectory requires the system to be controllable. You must be able to command the robot through motions that sufficiently excite all its dynamic modes to populate the regressor matrix meaningfully.

The basic formulation is extended to handle real-world complexities:

• Friction Modeling: Separate regressor columns are added for viscous and Coulomb friction parameters at each joint.
• Accounting for Unmodeled Dynamics: When residual errors persist after linear least-squares, it indicates unmodeled dynamics. Techniques like residual modeling with a neural network can be added to the regressor framework.
• Grey-Box Identification: The dynamic regressor represents the known physics (white-box). Its combination with data-driven models to capture residuals is a classic grey-box identification approach.
• Numerical Conditioning: Calculating q̈ from noisy q data is challenging. Alternatives include using the filtered regressor or instrumental variable methods to improve estimation robustness.

The overarching process of constructing mathematical models of dynamic systems from measured input-output data. A dynamic regressor is a key formulation derived within this process, enabling the linear estimation of parameters like mass and friction from sensor data.

• Core Objective: To characterize a system's behavior and estimate its unknown parameters.
• Pipeline Stages: Includes experiment design, data collection (using excitation trajectories), model selection, parameter estimation, and validation.
• Application: Directly enables the parameter calibration of physics simulators for sim-to-real transfer.

The specific process of inferring the unknown constant values within a system's mathematical model from observed data. Using a dynamic regressor transforms this into a linear least-squares problem.

• Target Parameters: Inertia tensors, link masses, viscous friction coefficients, and motor constants.
• Mathematical Relation: Y = Φ(q, q̇, q̈) * π, where Y is torque, Φ is the dynamic regressor, and π is the parameter vector.
• Contrast with State Estimation: Estimates constant model parameters, not the time-varying internal state of the system.

A deliberately designed sequence of robot motions or control inputs that is sufficiently rich to enable accurate system identification. It ensures the dynamic regressor matrix is well-conditioned for inversion.

• Purpose: To achieve persistent excitation, stimulating all dynamic modes of the system.
• Design Criteria: Often optimized to minimize the condition number of the regressor matrix or to maximize information gain.
• Example: A robotic arm following a sum-of-sinusoids joint trajectory to excite inertial and Coriolis forces.

The degree to which a simulation model accurately replicates the real-world system. Parameter calibration via a dynamic regressor is a primary method for improving physics-based fidelity.

• Quantification: Measured using fidelity metrics like mean squared error of joint torques or trajectory deviations.
• Limits: Even perfect parameter identification cannot overcome unmodeled dynamics or simulation bias.
• Outcome: High-fidelity models reduce the reality gap and transfer error for simulation-trained policies.

The performance discrepancy between a policy trained in simulation and its performance on real hardware. Inaccurate dynamic parameters, identified via an imperfect dynamic regressor, are a fundamental cause.

• Primary Contributors: Simulation bias, unmodeled dynamics, model uncertainty, and domain gap in sensory perception.
• Bridging Techniques: Includes domain randomization, system identification, and residual modeling.
• Measurement: Quantified as transfer error on task-specific metrics (e.g., success rate drop).

A hybrid modeling approach that combines first-principles physics with data-driven learning. The dynamic regressor provides the physics-based (white-box) structure, while data estimates the parameters.

• Structure: Torque = Φ(q, q̇, q̈) * π (Physics) + f_θ(q, q̇) (Neural Network).
• Advantage: More data-efficient and interpretable than pure black-box models, while compensating for residual effects.
• Use Case: Correcting for unmodeled dynamics like complex friction or flexure not captured in the rigid-body regressor.