Observability

A linear time-invariant system is defined as observable if, for any possible initial state vector x(0), knowledge of the input u(t) and output y(t) over a finite time interval [0, T] is sufficient to uniquely determine x(0). This is formally tested using the Observability Matrix, constructed from the system's A (state transition) and C (output) matrices. If this matrix has full column rank, the system is observable. For nonlinear systems, observability is assessed locally using the Observability Rank Condition, which involves Lie derivatives of the output function.

Observability and controllability are mathematical duals, forming the foundation of modern control theory (Kalman's duality principle).

• Controllability: Determines if inputs can drive the system to any desired state.
• Observability: Determines if the current state can be deduced from outputs. A system can be controllable but not observable (e.g., a hidden internal state), or observable but not controllable. For a system to be fully managed—both estimated and controlled—it typically requires both properties. This duality is exploited in observer design, such as the Luenberger observer or Kalman filter.

When a system is observable, we can construct an observer—a dynamic model that uses the system's measured outputs to estimate its internal states. Key observer types include:

• Luenberger Observer: A deterministic state estimator for linear systems.
• Kalman Filter: An optimal recursive estimator for linear systems with Gaussian noise, minimizing the mean squared error.
• Extended Kalman Filter (EKF): A nonlinear version that linearizes the system around the current estimate.
• Nonlinear Observers: Designed for systems where linearization fails, using techniques like sliding mode or high-gain observers. The observer's convergence rate is a critical design parameter.

Observability is a critical assumption in system identification. To accurately estimate a system's unknown parameters (e.g., inertia, friction), the experimental data must contain enough information. This is enforced through persistent excitation of the inputs.

• Identifiability: A related concept; a parameter is identifiable if it can be uniquely determined from input-output data. Lack of observability can lead to non-identifiable parameters.
• Experiment Design: Identification protocols must ensure the collected data makes the system states (and thus parameters) observable. This often involves designing specific excitation trajectories that probe all dynamic modes.

In real-world robotics and simulation, perfect observability is often an ideal. Challenges include:

• Sensor Limitations: Limited or noisy measurements (e.g., only position, not velocity) can reduce observability.
• Nonlinearities: A system may be observable in one region of state space but not in another.
• Unobservable Subspaces: Certain state combinations may not affect the output, making them impossible to distinguish. Engineers must analyze observability to determine optimal sensor placement and whether a state estimator (like a Kalman filter) will work reliably before deployment.

The Kalman Filter is the quintessential algorithm leveraging observability. It provides a probabilistic framework for state estimation in noisy, dynamic systems. Its operation assumes the underlying system is observable.

• Prediction Step: Uses the system model to project the state estimate forward.
• Update (Correction) Step: Uses the new sensor measurement to correct the prediction. The filter's innovation (measurement residual) is the difference between the actual and predicted measurement. The Kalman Gain optimally weights the confidence between the model prediction and the sensor reading. If the system is unobservable, the filter's error covariance for certain states will not decrease.

Observability dictates the minimum viable sensor suite for a robotic system. Engineers must strategically place sensors to make critical internal states inferable. For example:

• A mobile robot's global position is often unobservable using only wheel encoders due to slip; it requires an external reference like GPS, LiDAR-based localization, or visual odometry.
• A robotic arm's joint torque may be unobservable without a dedicated load cell, requiring estimation from motor current and a dynamic model. Poor sensor placement that leads to unobservable states results in systems that cannot be accurately identified or controlled in reality, creating a fundamental reality gap.

To build an accurate simulation model via system identification, the robot must execute motions that make its parameters observable. This involves designing excitation trajectories that satisfy persistent excitation conditions.

• Rich Trajectories: Simple, slow motions may not excite nonlinear dynamics like friction or backlash. Identification requires complex, sweeping motions that stress all dynamic modes.
• Practical Constraint: In the real world, excitation trajectories must respect physical limits (torque, velocity) and safety constraints, which can conflict with ideal identification signals. This trade-off is a core challenge in sim-to-real transfer.

Observability determines which state estimation techniques (e.g., Kalman Filters, observers) are viable. For partially observable systems, engineers deploy sensor fusion to reconstruct the full state.

• Fusing Complementary Sensors: Combining an IMU (observes orientation and acceleration) with a camera (observes position relative to landmarks) can make a drone's full 6D pose observable.
• Simulation's Role: Observability analysis is first performed in simulation to design and test estimation pipelines before costly real-world deployment. An unobservable configuration in simulation will fail in reality.

If a physical parameter is not observable given the available real-world sensors, it cannot be accurately calibrated. This leads to uncalibratable simulation bias.

• Example: The exact coefficient of friction between a specific rubber wheel and a polished concrete floor may be unobservable without a direct force-torque sensor at the contact point. A simulator can only use an approximate, generic value.
• Implication: This inherent uncertainty must be accounted for via robust control or domain randomization during policy training, as the simulation can never be perfectly matched to reality.

When a policy performs well in simulation but fails on hardware, lack of observability is a key root cause to investigate.

• Failure Mode: The policy may rely on a state that is perfectly known in simulation (e.g., exact contact forces) but is unobserved or poorly estimated in the real world.
• Mitigation: The solution involves either retraining the policy using only observable states (making it more robust) or augmenting the real system with additional sensors to restore observability, impacting system cost and complexity.

The construction of a high-fidelity digital twin is guided by observability. Engineering effort is focused on modeling dynamics that are both impactful and observable.

• High-Priority Modeling: Dynamics corresponding to highly observable states (e.g., arm joint angles from encoders) must be modeled with high precision.
• Lower-Priority Modeling: Dynamics that are fundamentally unobservable in the real system may be simulated with simpler, computationally cheaper models, as their ground-truth values can never be validated. This enables efficient use of simulation compute resources.

System identification is the process of constructing mathematical models of dynamic systems from measured input-output data. The goal is to characterize system behavior and estimate unknown parameters. A standard system ID pipeline includes:

• Designing excitation trajectories
• Collecting sensor data
• Selecting a model structure (e.g., grey-box)
• Performing parameter estimation
• Quantitative validation against held-out data

Controllability is a dual concept to observability. It measures whether an external input (e.g., actuator commands) can move a system's internal state from any initial point to any desired final state within a finite time. For a linear system, the pair (A, B) must be controllable. In robotics, a system must be both controllable and observable for effective closed-loop control and accurate state estimation.

State estimation is the practical application of observability. It is the process of inferring a system's complete internal state (e.g., position, velocity, bias) from a sequence of noisy, partial sensor measurements. Common algorithms include:

• Kalman Filters (for linear systems)
• Extended Kalman Filters (EKF) and Unscented Kalman Filters (UKF) (for nonlinear systems)
• Particle Filters (for highly nonlinear, non-Gaussian systems) These filters fuse sensor data with a dynamic model to produce optimal estimates.

While state estimation infers time-varying states, parameter estimation infers constant or slowly varying parameters of the system model itself. Examples include link mass, inertia tensors, and friction coefficients. Methods include:

• Least-squares regression on a dynamic regressor
• Bayesian calibration, which provides uncertainty bounds
• Data-driven calibration using optimization Accurate parameter estimation reduces model uncertainty and simulation bias.

Persistent Excitation (PE) is a condition required for successful system identification. An input signal is persistently exciting if it continuously provides enough energy across all frequencies of interest to uniquely identify all system parameters. Without PE, the parameter estimation problem becomes ill-posed. Engineers design specific excitation trajectories (e.g., chirp signals, random motions) to guarantee PE during data collection for system identification.

Model fidelity is the degree to which a simulation matches reality. Observability is critical for assessing it. The validation process involves:

1. Ground truth alignment: Synchronizing sim and real data.
2. Calculating fidelity metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE).
3. Analyzing calibration error and transfer error. Low-fidelity models suffer from unmodeled dynamics and a wide reality gap, making sim-to-real transfer challenging.