Constraint-Aware Discrete-Time PID Gain Optimization for Robotic Joint Control Under Actuator Saturation

Technical Summary: Constraint-Aware Discrete-Time PID Gain Optimization for Robotic Joint Control Under Actuator Saturation

Problem Statement
While Proportional-Integral-Derivative (PID) controllers are the standard for ensuring asymptotic stability in linear time-invariant (LTI) systems, their practical application in autonomous robotics is frequently compromised by hardware constraints. Theoretical models often assume continuous-time signal processing and infinite control authority. However, physical robotic joints operate within strict voltage or torque limits and utilize digital processors that execute control laws in discrete time steps. This discrepancy creates a performance gap: when a controller demands an input exceeding the actuator's physical capacity ($|u(t)| > u_{sat}$), the system enters a saturation regime. This introduces non-linearities, specifically integral windup, which degrade tracking fidelity (measured by Integral of Absolute Error, IAE) and increase settling times. Furthermore, real embedded loops suffer from non-idealities such as computation delays, sensor noise, and encoder quantization, which interact with discretization and saturation to destabilize loops that appear stable in continuous-time analysis.

Methodology
The authors present a systematic characterization of a first-order LTI robotic joint under discrete PID control, moving from theoretical derivation to robust, implementation-aware tuning.

1. Discrete-Time Modeling and Stability Analysis:

   • The study models the plant using a first-order LTI equation and discretizes it using both Forward-Euler and exact Zero-Order-Hold (ZOH) methods.
   • Stability Regions: Using the Jury criterion, the authors derive closed-form stability regions for PI control. These regions provide explicit, sampling-dependent constraints on gains ($K_p, K_i$) for both Euler and ZOH discretizations, allowing for "safe" gain selection prior to simulation.
   • Anti-Windup: A discrete back-calculation anti-windup update is implemented to handle saturation. The integral state is updated based on the difference between the commanded and saturated effort, parameterized by a correction gain $K_{aw}$.

2. Performance Metrics:

   • Evaluation relies on standard step-response metrics: Percent Overshoot (%OS), Rise Time ($t_r$), Settling Time ($t_s$), Terminal Steady-State Error ($e_{ss}$), and IAE.
   • A specific "saturation duty" metric is introduced to quantify the fraction of time the actuator remains saturated.

3. Robust Optimization Workflow:

   • Uncertainty Modeling: To emulate real-world conditions, the authors evaluate gains across a randomized family of models including uncertain parameters ($\tau, K$), input delays, measurement noise, quantization, and tighter saturation limits.
   • Hybrid-Certified Bayesian Optimization (HC-SBO): The paper proposes a tuning workflow that optimizes a robust objective function $J$. This objective aggregates median IAE with soft penalties for overshoot and saturation duty.
   • Safety Filtering: To prevent unsafe exploration, the workflow employs a "hybrid certification" filter:
     • Analytic Feasibility: Candidates are restricted to the ZOH Jury-feasible PI region.
     • Behavioral Certification: A short-horizon check on a second-order actuator model rejects candidates causing extreme overshoot or sustained saturation before expensive robust evaluation.

Key Contributions
The paper identifies three primary contributions aimed at bridging the gap between theoretical control and robotics implementation:

1. Closed-Form PI Stability Regions: Derivation of explicit gain-and-sampling-dependent stability constraints for PI control under both Forward-Euler and exact ZOH discretizations. This enables the pruning of unstable search spaces before simulation.
2. Saturation-Dominant Evaluation: An evaluation of discrete back-calculation anti-windup in regimes where saturation is frequent (a condition not captured by small-step, low-demand baselines), demonstrating how it reshapes recovery dynamics and overshoot.
3. Robust Constraint-Aware Tuning: A Bayesian optimization workflow that optimizes IAE while enforcing stability and soft constraints on overshoot and saturation duty across a randomized model family.

Experimental Results

• Baseline Sweeps: For a delay-free first-order plant, P-control reduces steady-state error but cannot eliminate it; PI control eliminates steady-state error, with $K_i$ dominating IAE reduction. Derivative action ($K_d$) showed minimal impact on the simple first-order plant but became beneficial for damping when delay and second-order dynamics were introduced.
• Saturation Dominance: In scenarios forcing early transient saturation, the inclusion of anti-windup significantly improved recovery times and reduced overshoot compared to standard integrators.
• Robustness and Efficiency:
  • Under a randomized model family (uncertainty, delay, noise, quantization), robust-oriented tuning improved the median IAE from 0.687 (manual baseline) to 0.470, while keeping median overshoot below 2%.
  • The hybrid certification screen rejected 11.6% of randomly sampled gains within bounds before full robust evaluation, improving sample efficiency without requiring hardware experiments.
  • The workflow successfully balanced tracking performance with saturation exposure, preventing the "chattering" or aggressive behavior often seen when optimizing purely for tracking error.

Significance and Claims
The paper claims to provide a grounded, implementation-aware framework for PID tuning that acknowledges the non-linearities of discrete-time execution and actuator limits. By deriving analytic stability guards and coupling them with a safety-aware Bayesian optimization workflow, the authors demonstrate that it is possible to achieve robust performance in simulation without hardware access. The work emphasizes that sampling rates, saturation, and small delays fundamentally alter the safe gain set and closed-loop behavior, necessitating a shift from continuous-time heuristics to discrete-time, constraint-aware design. The proposed "hybrid-certified" approach offers a practical method for rejecting unsafe candidates early, thereby making data-efficient tuning feasible for complex robotic joints where direct hardware iteration is impractical.