Controllability

A linear time-invariant system is defined as state-controllable if, for any initial state x(0) and any desired final state x_f, there exists a finite time T and a control input u(t) that transfers the state from x(0) to x_f in time T. For a system defined by ẋ = Ax + Bu, this is mathematically determined by the rank of the controllability matrix: [B, AB, A²B, ..., A^(n-1)B]. The system is fully controllable if and only if this matrix has full rank (equal to the state dimension n).

These are related but distinct properties. Controllability is a stronger condition, requiring the ability to reach any state. Stabilizability is a weaker condition, requiring only that all unstable modes of the system are controllable. A system can be stabilizable but not fully controllable if it has stable, uncontrollable modes. This distinction is crucial for practical control design, where stabilizing an unstable system to an equilibrium (like a balancing robot) is often sufficient, even if full state-to-state maneuverability is not possible.

For analyzing controllability over a finite time horizon and the energy required for state transfer, engineers use the Controllability Gramian. For a system ẋ = Ax + Bu, the Gramian over time [0, T] is defined as the integral: W_c(T) = ∫_0^T e^(Aτ) B B^T e^(A^T τ) dτ. This matrix is positive definite if and only if the system is controllable. Its eigenvalues indicate the "ease" of controlling the system in different directions of the state space; small eigenvalues correspond to states that require large control energy to reach.

While state controllability concerns the internal state x, output controllability concerns the system's measurable outputs y = Cx. A system is output controllable if there exists an input sequence that can transfer the output from any initial value to any final value within finite time. The condition depends on the rank of the matrix [CB, CAB, CA²B, ..., CA^(n-1)B]. A system can be output controllable even if it is not state controllable, which is common when sensors cannot observe the full internal state.

In robotics, underactuation—having fewer independent control inputs than degrees of freedom—directly challenges controllability. Classic examples include:

• Inverted Pendulum (Cart-Pole): A single horizontal actuator controls both the cart's position and the pole's angle.
• Quadrotor Drone: Four rotors control six degrees of freedom (position and orientation). These systems are often nonlinear and may only be controllable in certain regions of their state space or through dynamic maneuvers. Analyzing their controllability requires nonlinear methods like Lie algebra rank conditions.

Controllability analysis informs fundamental design and simulation decisions:

• Actuator Placement: Determines if motors/joints are positioned to exert authority over all critical system modes.
• Trajectory Planning: Validates whether a desired motion path is physically achievable by the system's actuation.
• Simulation Fidelity: A high-fidelity simulation model must preserve the controllability properties of the real robot. Incorrect modeling of actuator limits or dynamics can make a simulated robot appear controllable in ways the real hardware is not, creating a reality gap.
• Failure Analysis: Identifies which system failures (e.g., a stuck joint) render the robot uncontrollable.

Observability is the dual property to controllability, measuring how well the internal states of a dynamic system (e.g., joint velocities, internal forces) can be inferred from knowledge of its external outputs (sensor measurements) over a finite time horizon. A system is observable if its complete state vector can be uniquely determined from the output history.

• Critical for State Estimation: Observability is a prerequisite for algorithms like Kalman filters to accurately reconstruct a system's full state from noisy sensor data.
• Joint Analysis: For linear time-invariant systems, the Kalman rank condition tests for complete observability, analogous to the controllability test.
• Real-World Impact: In robotics, an unobservable system means certain internal conditions are indistinguishable, leading to estimation failures and potential control instability.

System identification is the experimental process of constructing mathematical models of dynamic systems from measured input-output data. It characterizes both the model structure and its unknown parameters (e.g., mass, inertia, friction).

• Enables Accurate Simulation: The resulting models are essential for high-fidelity simulators used in Sim-to-Real Transfer Learning.
• Two Main Approaches: Grey-box identification combines known physics with data-learned residuals, while black-box identification uses purely data-driven models like neural networks.
• Pipeline: A standard system ID pipeline involves designing excitation trajectories, collecting data, estimating parameters, and performing quantitative validation.

These are the core computations for simulating and controlling motion.

• Forward Dynamics: Calculates a system's acceleration and subsequent motion trajectory, given its current state and the applied forces/torques. This is the fundamental computation performed by a physics engine during simulation.
• Inverse Dynamics: Calculates the forces or torques required at the actuators to produce a desired acceleration or trajectory, given the current state and dynamic model. This is critical for model-based control and calculating the dynamic regressor for parameter estimation.
• Relationship: Accurate inverse dynamics relies on a high-fidelity model derived from system identification. Errors in these calculations directly impact controllability by providing incorrect control inputs.

State estimation is the process of inferring the internal, often unmeasured, state variables of a dynamic system (e.g., position, velocity, bias) from a sequence of noisy sensor observations and a system model.

• Depends on Observability: A system must be observable for its state to be fully estimable.
• Common Algorithms: Includes Kalman filters (linear), Extended Kalman Filters (nonlinear), and particle filters.
• Critical for Control: Most advanced controllers (e.g., LQR, MPC) require full state feedback. Since not all states are directly measured (e.g., velocity from a position encoder), accurate state estimation is essential for realizing the system's theoretical controllability in practice.

Persistent excitation is a property of an input signal to a dynamic system, ensuring the signal provides sufficient stimulation over time to allow for the consistent identification of all the system's parameters and dynamic modes.

• Prerequisite for Reliable ID: Without persistent excitation, the system identification process may yield inaccurate or non-unique parameter estimates.
• Designing Trajectories: Engineers create specific excitation trajectories (e.g., chirp signals, random motions) that guarantee persistent excitation for the target parameters.
• Link to Controllability: A system must be controllable to execute the rich excitation trajectories needed for identification. Furthermore, accurately identified models improve model-based control, effectively enhancing practical controllability.

Model uncertainty quantifies the lack of perfect knowledge about a system's true dynamics, arising from simplifications, unmodeled dynamics, or inaccurate parameter values.

• Sources: Includes simulation bias, calibration error, and approximations in contact mechanics.
• Impacts Controllability: High model uncertainty degrades the performance of model-based controllers, as the computed control laws are based on an inaccurate representation of the system. This can make a theoretically controllable system difficult to control in practice.
• Mitigation: Techniques like robust control, adaptive control, and residual modeling are used to maintain performance despite uncertainty.