Closed-Loop Simulation Architecture

• Closed-loop simulation is a system design paradigm featuring bi-directional interactions between agents and simulated environments that enable continuous feedback and dynamic adaptation.
• It integrates sensor simulation, control algorithms, and state-update modules to mimic real-world uncertainties and support applications in autonomous driving, robotics, and biomedical modeling.
• Advanced frameworks employ mathematical modeling, optimization techniques, and real-time feedback correction to enhance system robustness and performance evaluation.

A closed-loop simulation architecture is a system design paradigm in which simulated agents or controllers interact bi-directionally with a simulated environment, enabling real-time feedback, dynamic behavior, and high-fidelity evaluation. Unlike open-loop approaches—where agent actions do not influence the subsequent state of the environment—closed-loop systems continuously update the simulated world based on agent actions, sensor feedback, and evolving environmental conditions, thereby replicating the feedback dynamics and uncertainties characteristic of physical or operational systems. This methodology has become foundational across domains such as autonomous driving, robotics, systems biology, cardiovascular modeling, and complex cyber-physical infrastructure.

1. Fundamental Principles and Taxonomy

Closed-loop simulation architectures are characterized by the cyclical exchange of information between agents (software, physical, or hybrid) and their environment. At each simulation step, the system evaluates sensor data (real or synthetic), computes actions or control commands via a policy or controller, updates the state of the environment and agents accordingly, and then propagates sensory updates—realizing an interleaved perception–decision–action loop (Sarva et al., 2023, Yan et al., 2024, Weidel et al., 2016, Piersanti et al., 2021, Coulombe et al., 23 Jun 2025).

Closed-loop frameworks manifest in multiple paradigms:

• Agent-centric architectures, where agents receive sensory input (images, point clouds, state vectors), process this via a policy or model, and affect the world through simulated actuators. This is standard in autonomous driving (Sarva et al., 2023, Yan et al., 2024, Lu et al., 1 Aug 2025).
• System-level feedback architectures, in which coupled ODE/PDE models for physical, biological, or engineered systems exchange state variables, forces, or signals through feedback controllers (PI, PID, DFACS, etc.) (Piersanti et al., 2021, Heisenberg et al., 2023, Coulombe et al., 23 Jun 2025).
• Hybrid hardware-in-the-loop (HIL) systems, wherein software simulation is tightly integrated with real-world sensors, actuators, or communication links (Nayak et al., 2018, Hoque et al., 2018).
• Adversarial and optimization-based closed-loop simulation, where environment or agent parameters are dynamically perturbed in feedback to stress-test system robustness or discover failure modes (Sarva et al., 2023, Chang et al., 2023).

2. Architectural Components and Data Flows

Canonical closed-loop simulation pipelines combine interconnected modules with explicit control, perception, and environmental propagation blocks. Core architectural elements include:

• Scene/Sensor Simulation: High-fidelity rendering of sensory data (e.g., LiDAR point clouds, images, EGM signals) from the current world state (Sarva et al., 2023, Yan et al., 2024, Lydon et al., 2 May 2025).
• Agent/Controller: Policy networks, planners, finite-state machines, RL agents, or differential equation solvers compute actions or control signals based on sensory input (Yan et al., 2024, Coulombe et al., 23 Jun 2025, Piersanti et al., 2021).
• Actuation & State Update: The simulated system evolves according to physics models, vehicle dynamics, or physiological ODE/PDEs using the agent's action (Yan et al., 2024, Heisenberg et al., 2023).
• Feedback/Assimilation: Observation or state assimilation modules integrate new real or simulated measurements, correcting or informing the simulation state to mitigate drift (Schenck et al., 2017).
• Black-Box or Adversarial Optimizers: In some frameworks, a search loop proposes worst-case parameters or adversarial scenarios, guided by performance metrics evaluated in simulation (Sarva et al., 2023, Chang et al., 2023).

The data flow is always cyclic: state and observations → agent decision or controller → environment update → new state and observations.

A table summarizing representative architectures:

Application Domain	Closed-Loop Modules	System-Specific Features
Autonomous Driving	Sensor sim, autonomy stack, world sim	Adversarial shape search, LiDAR, BO
Cardiac Electromech.	3D EP, mechanics, 0D circulation	Volume-constrained PDE/ODE coupling
Power Electronics	PINN surrogate, PI controller	Physics-constrained loss, sequence net
Neurorobotics	ROS–MUSIC bridge, SNN/robot sim	Real-time spike-to-actuator feedback

3. Mathematical Formulations and Optimization in the Loop

Closed-loop architectures leverage explicit mathematical models to ensure proper propagation of feedback and to enable quantitative evaluation and optimization:

• State-Space or PDE/ODE Models: Dynamics are defined by equations such as $\dot{x} = f(x,u)$ (robotics, electronics) or more complex PDE systems for biomechanical tissue or fluids (Piersanti et al., 2021, Heisenberg et al., 2023).
• Cost or Reward Accumulation: Many architectures define cumulative performance metrics (objective functions) over time. For adversarial testing, costs may combine penalties on perception $\ell_{\mathrm{det}}$, prediction $\ell_{\mathrm{pred}}$, planning $c_{\mathrm{plan}}$ (Sarva et al., 2023):

$$J(\mathbf z) = \sum_{t=1}^T \mathcal{C}_t(\mathcal{S}_t, \mathcal{F}(\widetilde\psi(\mathcal{S}_t, \mathcal{G}(\mathbf z), \mathcal{E}_t)))$$

These objectives are optimized via Bayesian Optimization or other black-box methods in the outer loop (Sarva et al., 2023).

• Autoregressive or Recurrent Inference: Time-series surrogate models (BiLSTM, Transformers) are used to propagate dynamics in response to feedback, particularly for multi-agent or power electronics domains (Coulombe et al., 23 Jun 2025, Lu et al., 1 Aug 2025).
• Feedback Correction and Data Assimilation: Methods like MAP filtering or force-based assimilation correct simulator state using real or virtual sensor observations, e.g., pulling simulated particles to match segmentation masks in fluid simulation (Schenck et al., 2017).
• Metric Constrained Optimization: Closed-loop benchmarks enforce safety, comfort, or task-completion metrics, guiding the agent's adaptation or exposing failure modes in a statistically robust setting (Zhang et al., 4 Aug 2025, Sarva et al., 2023).

4. Domain-Specific Implementations and Exemplars

Closed-loop simulation is instantiated across a range of scientific and engineering applications:

Autonomous Driving:

Adv3D simulates a full autonomy stack (perception, prediction, planning), with synthetic LiDAR sweeps generated from digital-twin reconstructions and adversarially perturbed actor geometries, closing the loop via interactive feedback and black-box optimization (Sarva et al., 2023). Benchmarks such as DriveE2E and DriveArena close the feedback from world state, to agent, to environment, ruling out open-loop abstraction mismatches (Yu et al., 28 Sep 2025, Yang et al., 2024). SAFE-SIM injects adversarial agent objectives into a diffusion-controlled multi-agent loop, maximizing collision likelihoods while enforcing physical realism (Chang et al., 2023).

Physiology and Biomechanics:

3D–0D cardiac models link high-dimensional PDE-based heart mechanics with a zero-dimensional ODE model of the full circulatory system. Volume and pressure variables are coupled as feedback constraints, forming a genuinely closed loop that matches experimental physiological markers and admits perturbation analysis (Piersanti et al., 2021). SimICD tightly integrates an electrophysiology solver with device decision logic, using asynchronous file-driven interfaces and checkpointing for feedback therapy injection (Lydon et al., 2 May 2025).

Power Electronics:

Physics-informed BiLSTM models simulate the time-domain evolution of dc–dc converters, closing the loop with a PI controller, with PINN terms in the loss to satisfy power-balance and ensure physical plausibility (Coulombe et al., 23 Jun 2025).

Neurorobotics:

Middleware bridges, e.g., ROS–MUSIC, enable real-time feedback between spiking neural networks and robotic worlds, synchronizing sensori-motor signals at millisecond latencies via continuous- and spike-based messaging (Weidel et al., 2016).

Parallel and Distributed Systems:

IDCVS and related frameworks couple traffic simulators (SUMO) with network simulators (OMNET++) under tight synchronization constraints, enabling city-scale, real-time closed-loop operation with hardware-in-the-loop sensor fusion (Hoque et al., 2018).

5. Design Trade-offs, Synchronization, and Scalability

Closed-loop simulation imposes nontrivial requirements for computational, architectural, and synchronization fidelity:

• Physical Realism vs. Efficiency: High-fidelity perception and physics increase compute costs (e.g., hours of GPU time for 100-query Bayesian optimization in Adv3D), requiring pragmatic trade-offs between simulation fidelity and throughput (Sarva et al., 2023).
• Real-time Synchronization: Co-simulation frameworks (e.g., LISANode, ROS–MUSIC) maintain alignment between feedback loops using MPI barriers, MUSIC clocks, or time-stepped discretizations, ensuring wall-clock or simulation-clock correctness (Weidel et al., 2016, Heisenberg et al., 2023).
• Domain Decomposition and Parallelism: Large-scale transportation or city simulators use hybrid parallelism (partitioned nodes each running coupled SUMO–OMNET++ pairs, OpenMP within node, MPI across nodes), managing cross-boundary synchronization with round-robin or conservative lookahead (Hoque et al., 2018).
• Hardware-in-the-Loop and Fault Tolerance: By integrating real radios, sensors, and controllers in simulation (DSRC units, physical controllers), distributed architectures enable realistic communication delays, packet loss, and failure recovery (Nayak et al., 2018).
• Modularity and Adaptability: State-of-the-art frameworks are modular, allowing agents, world models, and optimizers to be interchanged or upgraded independently with minimal system-level impact (Yang et al., 2024, Jiang et al., 23 Oct 2025).

6. Empirical Outcomes and Evaluation Metrics

Closed-loop simulation systems are evaluated using a variety of task- and system-specific metrics. Benchmark results include:

• Degradation in detection/perception: Closed-loop adversarial shape search can degrade 3D detection AP by 8–10 points and increase prediction error (ADE) by ∼0.2 m compared to open-loop or baseline configurations (Sarva et al., 2023).
• Comfort and safety penalties: Planning comfort costs (jerk, lateral acceleration) often double with closed-loop adversarial perturbation (Sarva et al., 2023).
• Task and behavior benchmarks: Success Rate, Driving Score, collision and route-completion rates are systematically lower in closed-loop than open-loop, reflecting the increased difficulty and interaction realism (Zhang et al., 4 Aug 2025, Yu et al., 28 Sep 2025).
• Simulator consistency and consistency-in-transfer: Reduced RMSE and variance in surrogate-based power electronics models, and strong correspondence between sim and real outcomes in robotics manipulation and cardiovascular simulation (Coulombe et al., 23 Jun 2025, Piersanti et al., 2021, Jiang et al., 23 Oct 2025).
• System scalability: Parallel/distributed closed-loop simulators achieve substantial speed-up (up to 12–20x) on urban networks, with communication overhead growing with the number of partition edges (Hoque et al., 2018).

7. Theoretical and Practical Significance

Closed-loop simulation architectures represent the state-of-the-art methodology for evaluating interactive, feedback-sensitive, and safety-critical systems. Recent frameworks provide:

• Quantitative tools for exposing rare or long-tail failures by adversarially stress-testing policies or controllers in feedback (Sarva et al., 2023, Chang et al., 2023).
• Infrastructure for integration of real-time sensor and actuator data, leading to hybrid simulation/real-world digital twins (Nayak et al., 2018, Jiang et al., 23 Oct 2025).
• Scientifically robust settings for validating physiologically or operationally complex control loops (e.g., cardiac therapy, drag-free spacecraft, power electronics) (Piersanti et al., 2021, Heisenberg et al., 2023).
• Modular, scalable software platforms that accommodate future advances in policy learning, sensor simulation, and optimization-based scenario synthesis (Yang et al., 2024, Lu et al., 1 Aug 2025).

The architecture's generality and extensibility ensure its ongoing impact across computational science, robotics, cyber-physical systems, and simulation-based safety engineering.