AI Control Scientist: LLM-driven Agentic System for Automated Control Design

Motivation

• Increasing system complexity makes heuristic control tuning labor-intensive, time-consuming, and difficult to scale under couplings, uncertainties, and multiple constraints.Expert knowledge is also fragmented and difficult to convert into reusable automated workflows.Source: §I · p. 1

Challenges

• General-purpose LLMs often lack the grounded physical reasoning needed to synthesize controllers handling stability margins, saturation, coupling, and dynamic trade-offs.The challenge spans multiloop PID, constrained MPC, and nonlinear feedback compensation.Source: §I · pp. 1–2
• Ambiguous natural-language requirements do not directly specify the mathematical objectives and constraints required for industrial control design.Existing LLMs struggle to map descriptions such as avoiding overshoot or improving disturbance rejection into rigorous formulations.Source: §I · pp. 1–2
• Unreliable numerical reasoning makes direct LLM parameter optimization unsuitable for fine-grained tuning, leaving practitioners dependent on repeated simulation-based expert adjustment.The bottleneck includes PID gains, observer poles, and MPC weighting matrices.Source: §I · p. 2

Proposed system

• AICS uses a Task Modeling Agent, Controller Design Agent, and Parameter Tuning Agent to translate requirements into models, generate controller code, and refine parameters under closed-loop criteria.The agents collaborate iteratively to emulate the workflow of experienced control engineers.Source: §I · p. 2

Results

• Experiments show higher design success and optimization efficiency than automated baselines, with performance comparable to domain experts across representative control systems.The paper positions the workflow as a scalable and generalizable paradigm for AI-native control engineering.Source: §I · p. 2

Proposed system

• The system combines logic-guided controller generation with planner-solver optimization, using inferred system characteristics and external numerical solvers for precise tuning under practical constraints.The Design Blueprint guides code synthesis, while the LLM supplies search bounds and priors to numerical solvers.Source: §I · p. 2

LLMs for code generation

• LLM software-engineering systems have progressed from code completion toward autonomous program synthesis and iterative refinement.Closed-loop reflection and planning methods treat code generation as a nonlinear state-space search.Source: A. Large Language Models for Code… · p. 2
• Execution feedback and reinforcement-learning formulations improve functional code generation by aligning textual outputs with variable trajectories, compiler observations, and interactive environments.CodeRL+ and InterCode exemplify this execution-centered direction.Source: A. Large Language Models for Code… · p. 2

LLMs for control engineering

• Control-oriented code agents transform natural-language instructions into executable Python or C++ control programs, with later systems adding environmental feedback and persistent skill libraries.These approaches bridge high-level requirements and implementable control behavior.Source: B. Large Language Models for Automated… · p. 3
• ControlAgent and AgenticControl use specialized multi-agent workflows and computation tools to evaluate frequency-domain behavior and refine controller parameters across increasingly complex scenarios.AgenticControl extends the architecture to six agents and addresses nonlinear dynamics and parametric uncertainty.Source: B. Large Language Models for Automated… · p. 3

Framework

• Control design is formulated as a high-dimensional, constrained, non-convex optimization problem over a control law and its parameters.The objective incorporates system evolution and physical or performance constraints.Source: A. Framework · p. 3
• The pipeline decomposes global optimization into Semantic Formalization, Code Generation, and Parameter Tuning stages executed by three agents.This sequential workflow starts from natural-language instructions and ends with tuned controller parameters.Source: A. Framework · p. 3

Semantic Formalization

• The Task Modeling Agent removes semantic ambiguity and converts user objectives into deterministic system parameters, mathematical definitions, and performance constraints.It uses a translation knowledge base and control-standard verification and analysis tools.Source: A. Framework · p. 3

Code Generation

• The Controller Design Agent selects a control strategy and generates code through inference and synthesis, then iteratively refactors errors identified by syntax and control-theoretic verification.Critical system features such as time delays and coupling inform a Design Rationale.Source: A. Framework · p. 3

Parameter Tuning

• The Parameter Tuning Agent constructs a simulator-based closed loop and uses LLM reasoning to define rewards and search ranges for efficient numerical optimization.The tuned parameters are produced by searching within the resulting parameter space rather than by direct LLM arithmetic.Source: A. Framework · p. 3

Framework

• The complete process targets end-to-end automated control design while preserving theoretical correctness and engineering feasibility.The framework integrates natural-language input, controller synthesis, verification, and tuning.Source: A. Framework · p. 3

• The TMA maps unstructured input into a specification containing the plant dynamic model M and control constraints Cconstraint.Source: B. Task Modeling Agent (TMA) · p. 3
• Semantic formalization translates engineering descriptions into a deterministic system model and numerical performance boundaries.Source: B. Task Modeling Agent (TMA) · p. 4
• The constraint set separates physical execution boundaries Cphys from dynamic performance metrics Cperf.Source: B. Task Modeling Agent (TMA) · p. 4
• After formalization, verification converts LTI systems to standard matrix form and checks stability, physical feasibility, and constraint consistency.Source: B. Task Modeling Agent (TMA) · p. 4

• The CDA uses a system-characteristic-driven paradigm to tailor controller logic to the plant rather than relying on generic text-to-code generation.Source: C. Controller Design Agent (CDA) · p. 4
• It infers dynamical descriptors from Sspec and uses control knowledge to determine the control law π and a theoretical blueprint Blogic.Source: C. Controller Design Agent (CDA) · p. 4
• Conditioning code generation on Blogic makes the implementation a deterministic consequence of identified system characteristics rather than a stochastic guess.Source: C. Controller Design Agent (CDA) · p. 4
• Verification combines syntax checking with control-theory checks, including controllability criteria, to intercept invalid logic.Source: C. Controller Design Agent (CDA) · p. 4
• When verification fails, exception and theory-violation information is fed back to the LLM for controller regeneration until verification succeeds.Source: C. Controller Design Agent (CDA) · pp. 4–5

• The PTA addresses numerical hallucination by decoupling symbolic reasoning from numerical computation in a meta-optimization paradigm.Source: D. Parameter Tuning Agent (PTA) · p. 5
• Each meta-optimization configuration Φk contains a parameter search domain Ωk and a search reward function Rk(·).Source: D. Parameter Tuning Agent (PTA) · p. 5
• Using trial history, the LLM adjusts the reward function or search space instead of directly guessing control parameters.Source: D. Parameter Tuning Agent (PTA) · p. 5
• Under Φk, a numerical optimizer builds a surrogate objective model and samples Ωk by maximizing an acquisition function α(θ).Source: D. Parameter Tuning Agent (PTA) · p. 5
• Within a limited simulation budget, the optimizer obtains a parameter vector θ∗ that maximizes expected reward.Source: D. Parameter Tuning Agent (PTA) · p. 5
• After each closed-loop simulation, the system calculates performance metrics and updates historical memory for subsequent refinement.Source: D. Parameter Tuning Agent (PTA) · p. 5
• Algorithm 1 specifies the complete double-layer iterative tuning process driven by simulation.Source: D. Parameter Tuning Agent (PTA) · p. 5

• The study evaluates AICS through a series of comprehensive experiments.Source: §IV · p. 5
• Algorithm 1 takes a controller, target specifications, a maximum refinement-round count K, and a maximum simulation count N per round.Source: §IV · p. 5
• Each refinement round uses the LLM to design or refine an optimization configuration before numerical optimization proceeds.Source: §IV · p. 5
• The procedure compiles and simulates proposed parameters, debugging controller code when execution fails.Source: §IV · p. 5
• Successful simulations are evaluated, scored, and added to optimizer and refinement histories while retaining the best parameters.Source: §IV · p. 5
• The procedure returns when the target specification is met, otherwise it builds feedback for the next refinement round.Source: §IV · p. 5

• The benchmark covers stable, unstable, time-delay, and higher-order complex systems, with 50 independent test cases per category.Source: A. Experimental Setup · p. 5
• Pass Rate measures whether controllers satisfy control objectives, performance metrics, and constraints.Source: A. Experimental Setup · p. 5
• Number of Iterations measures the average refinement steps needed to reach the optimal parameter configuration.Source: A. Experimental Setup · p. 5
• Figure 3 tracks time-response evolution during tuning, while Figure 4 summarizes iteration-count distributions for successful trials.Source: Figure 3 · p. 6, Figure 4 · p. 6

• As complexity increases, vanilla Zero-shot and Few-shot methods fail completely in second-order unstable environments, with a 0% pass rate.Source: B. Main Results and Baseline Comparison · p. 6
• On highly nonlinear second-order unstable systems, AICS achieves a 90% pass rate using 2.44 iterations, versus ControlAgent’s 74% pass rate and PIDtune’s 12%.Source: B. Main Results and Baseline Comparison · p. 6
• On higher-order complex systems, AICS reaches a 74% pass rate, exceeding ControlAgent at 54%, Few-shot with feedback at 40%, and PIDtune at 50%.Source: B. Main Results and Baseline Comparison · p. 6
• ControlAgent reaches around 6 median iterations on challenging second-order unstable systems, with an upper quartile of 7 and outliers up to 9.Source: B. Main Results and Baseline Comparison · p. 7
• AICS more accurately uses historical feedback to identify performance flaws and rapidly execute corrective parameter updates than ControlAgent.Source: B. Main Results and Baseline Comparison · p. 7
• The ablation results indicate that removing numerical optimization causes severe pass-rate degradation, especially for higher-order and second-order unstable systems.Source: Figure 5 · p. 7

Ablation Design

• The ablations remove the Numerical Optimization and LLM Historical Feedback Guidance modules to test the necessity of these mechanisms.Source: C. Ablation Study on Multi-Agent… · p. 7

Numerical Optimization

• Removing numerical optimization causes severe pass-rate collapse on unstable and higher-order systems, with slower convergence and a lower final performance bound.The result supports decoupling symbolic reasoning from numerical computation in complex control design.Source: C. Ablation Study on Multi-Agent… · p. 7

Meta-Optimization Guidance

• Removing meta-optimization guidance slows early convergence and limits the final asymptotic pass rate to approximately 84%, versus over 90% for the full system.The guidance dynamically narrows the parameter search space using trial history.Source: C. Ablation Study on Multi-Agent… · p. 7

Model Scale and Reasoning

• Smaller models perform comparably on delay and first-order systems, while stronger reasoning and code synthesis improve performance on second-order unstable and higher-order systems.Source: D. Impact of Model Scale and Reasoning… · p. 8
• Qwen3-235B reaches a 74% pass rate on higher-order tasks, compared with 54% for the 8B model.Source: D. Impact of Model Scale and Reasoning… · p. 8

Workflow

• The execution proceeds through Semantic Formalization, Code Generation with error self-correction, and Simulation-driven Parameter Tuning.Source: E. Case Study: End-to-End Autonomous… · p. 8

Semantic Formalization

• The Task Modeling Agent maps linguistic constraints into deterministic parameters and establishes a mathematical feasible region through open-loop analysis.Example parameters include max_overshoot = 15.1 and max_settling_time = 5.3.Source: E. Case Study: End-to-End Autonomous… · p. 8

Code Generation

• The Controller Design Agent captures execution errors and refactors faulty controller logic using a control-theory verification engine without human intervention.Source: E. Case Study: End-to-End Autonomous… · p. 8

Parameter Tuning

• The Parameter Tuning Agent delegates numerical search to an optimizer while the LLM configures the reward function and search boundaries, producing Q = diag([50]) and R = diag().Source: E. Case Study: End-to-End Autonomous… · p. 8

Output Verification

• The verified controller achieves 4.8% overshoot, 0.1 s settling time, and 8 × 10^-6 steady-state error, all below the specified limits.Source: E. Case Study: End-to-End Autonomous… · p. 8