Human-in-the-Loop Control

• Human-in-the-Loop Control is an integrated approach that combines human expertise with machine feedback to enhance safety, adaptability, and performance in complex environments.
• It employs methods like blended autonomy, predictive feedback, and online learning, with applications in teleoperation, assistive robotics, and digital twin simulations.
• Research in HITL addresses challenges such as human model uncertainty, latency compensation, and adaptive authority allocation to ensure robust and scalable control.

Human-in-the-loop (HITL) control is an integrated systems paradigm in which human actions, preferences, state, or expertise are explicitly incorporated into feedback or decision loops alongside machines, sensors, and actuators. Unlike fully automated schemes, HITL architectures seek to exploit human cognitive, perceptual, or physical capacity to achieve robustness, safety, adaptability, or human-aligned objectives in complex control environments. HITL control frameworks occur in diverse domains, including teleoperation, collaborative robotics, exoskeletons, safety-critical infrastructure, supervisory systems, multi-robot teams, and adaptive autonomous agents. Methods range from explicit input blending and shared control to learning paradigms that adapt system parameters based on online human performance, preference modeling, or context-aware instruction.

1. Architectural Patterns and Algorithmic Principles

HITL control architectures are characterized by explicit pathways for human input, observation, or override, together with feedback mechanisms that maintain system-level guarantees in spite of unpredictable or suboptimal human actions. Typical structural motifs and core algorithmic principles include:

• Bilateral Teleoperation with Predictive Rendering: Systems such as the TactiMesh Teleoperator Interface integrate operator head pose and end-effector command streams into parallel visual and haptic pipelines. Visual feedback exploits incremental SLAM-based 3D reconstructions (via ORB-SLAM2 and CARV algorithms), while real-time force rendering closes the bilateral loop through a digital twin of the manipulator embedded in a physics simulator. Predictive models (e.g., Kalman filtering of SE(3) head-pose trajectories, end-effector dynamics in discrete state-space) compensate for network and estimation delays, enabling stable, low-latency feedback for immersive manipulation tasks (Akturk et al., 2024).
• Parametric Hybrid Control: In exoskeletons and assistive robotics, layered architectures divide functionality into low-level motion/torque control, mid-level adaptive/assist-as-needed (AAN) modules responsive to biomechanical signals, and outer HITL optimization loops. Online learning (Bayesian optimization, Gaussian process regression) tunes controller parameters to optimize cost functions that conjoin human participation and task performance, such as symmetry errors in gait, fostering an efficient, user-specific assistive strategy (Qian et al., 23 Mar 2025).
• Blended Autonomy and Arbitration: In settings such as lunar robotic assembly, the control system arbitrates between autonomous and human-supplied control commands by continuously modulating authority coefficients (α in [0,1]), depending on event-driven triggers (e.g., anomalous sensor readings, ambiguous perception). Error-detection, adaptive dynamic models, and digital-twin simulation pipelines collectively reduce operational risk and workload, while providing seamless operator intervention when environmental uncertainty or mission-critical ambiguities arise (Mishra et al., 15 Jul 2025).
• Mixed-Initiative and Safety-Constrained Fusion: For mobile robotics under Linear Temporal Logic (LTL) tasks, additive blending (u = u_r + κ(x)·u_h) enables humans to influence robot motion up to a smoothly varying blending factor κ(x) defined by proximity to unsafe or infeasible regions in product automaton state space. By analyzing reachable sets and trap states, controllers guarantee that “hard” LTL constraints are never violated, independently of human-driven proposals (Guo et al., 2018, Yu et al., 2023).
• Shared Control in Supervisory and Manual-Automatic Takeover: HITL architectures in advanced vehicles and air traffic control fuse real-time human and autonomous commands via explicit blending weights or artifact modules, often augmented by context-sensitive and regulated transition policies. Shared-control strategies are validated in simulator scenarios with abrupt handover conditions, using metrics such as time-to-collision and operator workload, highlighting the necessity of adaptive arbitration and individualized authority allocation (Zhou et al., 2021, Carvell et al., 7 Jan 2026).
• Stochastic and Behaviorally-Informed Human Modeling: In human-cyber-physical systems, learning human behavior models via Gaussian Mixture Models (GMMs), Markov chains, or fuzzy inference systems enables more accurate state prediction, less conservative reachability, and safety certification even when human “inputs” are poorly modeled, probabilistic, or systematically biased (Choi et al., 2022, Banerjee et al., 2024, Firouznia et al., 2018, Protte et al., 2020).

2. Predictive Feedback, Latency Compensation, and Stability Guarantees

A central challenge for HITL control is maintaining system stability and transparency in the presence of network latency, human sensorimotor delays, and unpredictable user dynamics.

• Predictive Visual and Haptic Feedback: In teleoperation, predictive rendering of future viewpoints and local closure of the haptic loop on digital twins offset the effect of variable network and pipeline delays. Linear motion models in SE(3) equipped with Kalman filtering predict operator pose; local physics simulation minimizes round-trip time for haptic feedback (<4 ms in (Akturk et al., 2024)).
• Passivity and Two-Port Network Analysis: The stability of closed-loop HITL systems, particularly in bilateral force-feedback, is established via passivity analysis of the overall transfer function and bounding network-induced delays according to the system’s cutoff frequency. Discrete-time passivity constraints (bounded stiffness and update rate) further prevent energy injection and resultant instability.
• Robustness to Human Action Variability: Weak control frameworks issue set-valued admissible control signals, so that any human-chosen action within the set maintains closed-loop input–output stability. By construction, the performance boundary can be explicitly tuned by controlling the geometric expansion of the action set, and learning algorithms can minimize additional human cost while preserving global guarantees (Inoue et al., 2018).
• Safety Under Hybrid Human–Machine Input: Constraint-based techniques—such as control barrier functions (CBF/CLBF) extended to handle both real-world controller and human in the plant—ensure safe set invariance even when human action plans are stochastic and data-driven. Markovian event models and fuzzy rules synthesize a plausible space of human-induced exogenous actions and derive controllers certifying probabilistic forward invariance (Banerjee et al., 2024).

3. Learning and Optimization with Human-in-the-Loop

Modern HITL controllers exploit online adaptation, preference optimization, and model updating driven by human input, feedback, or demonstration.

• Learning-by-Demonstration and Online Policy Adaptation: Architectures such as Hug-DRL blend autonomous policy learning with real-time human guidance and interventions. Reinforcement learning agents adjust their value and policy networks by incorporating both reward-prediction error and imitation loss on human-overridden steps, with adaptive trust weighting to balance human demonstrations and autonomous exploration. Policy updates leverage experience replay augmented with intervention flags, and convergence is catalyzed by the injection of meaningful, expert-corrected trajectories (Wu et al., 2021).
• Preference-Based Control Tuning and Optimization: Assist-as-needed exoskeletons combine iterative learning controllers with outer-loop Bayesian optimization, which adaptively tunes learning gains to minimize a composite cost function capturing spatial and temporal symmetry as well as control effort. Human biomechanical measurements inform the cost and adaptation law, driving personalized, subject-specific controller synthesis (Qian et al., 23 Mar 2025).
• Inverse Optimization for Supervisor-Driven Multi-Robot Coordination: In tasks where human supervisors may suggest action subsets non-aligned with expert-prescribed submodular combinatorial plans, algorithms for Inverse Submodular Maximization (ISM) compute the minimal parameter adjustments required for the baseline greedy algorithm to select the human-desired set, formulating this as a mixed-integer quadratic program and solving it efficiently via branch-and-bound over partial solution orderings (Shi et al., 2024).
• Online Fine-tuning with Language-In-The-Loop: In context-aware Model Predictive Control (InstructMPC), human operators or domain experts provide textual instructions about upcoming events, maintenance, or anomalies, which are embedded by a Language-to-Distribution (L2D) module (e.g., LLMs) to produce disturbance forecasts utilized in the MPC horizon. Parameters are then online-updated via loss functions designed to align task cost with prediction error, yielding closed-loop regret bounds O(√{T log T}) over finite time (Wu et al., 5 Dec 2025, Wu et al., 8 Apr 2025).

4. Human-in-the-Loop in Simulation, Planning, and Digital Twin Frameworks

HITL approaches increasingly leverage digital twins, immersive VR/AR, or simulator-based training for safe, efficient, and transparent system design and operation.

• Digital Twin-Enabled HITL: Environments such as Isaac Sim or Unity-based digital twins copy both physical context and robot state, allowing operators to intervene, test what-if scenarios, practice in the presence of rare faults, or validate control updates with full bidirectional data and outcome replay. Task refinement cycles involve real-time synchronization between digital and real systems, closing the gap between simulation and operational deployment (Mishra et al., 15 Jul 2025, Yigitbas et al., 2021).
• Procedural versus Declarative User Interaction: In VR-mediated HITL adaptive systems, procedural control grants maximal user authority by allowing fine-grained trajectory recording, whereas declarative (mixed-initiative) protocols enable users to specify only high-level goals, which are realized via automated planning and then simulated for verification. Both strategies involve synchronized digital twinning, with trade-offs between execution speed, transparency to the operator, and user effort (Yigitbas et al., 2021).
• Simulator-Based Evaluation and Human Subject Integration: Comprehensive design and validation of shared-control takeover regimes in driving, air traffic control, and assistive robotics employs high-fidelity simulators, enabling quantitative assessment (e.g., takeover times, collision proximity, workload indices, completion rates) and controlled exploration of diverse engagement and disengagement scenarios (Zhou et al., 2021, Carvell et al., 7 Jan 2026, Wang et al., 5 Mar 2025).

5. Applications and Empirical Validation

HITL control is applied in domains with a spectrum of uncertainty, autonomy, and safety requirements:

• Teleoperation and Hazardous Environment Manipulation: TactiMesh’s predictive, immersive visual-haptic teleoperation system enables rapid, precise manipulation in remote or hazardous settings (search and rescue, remote maintenance) with documented reductions in completion time and oscillatory tracking artifacts versus non-predictive baselines (Akturk et al., 2024).
• Assistive Devices and Wearable Robotics: Exoskeleton controllers using HITL optimization demonstrated significant improvements in gait symmetry, participation indices, and joint torque profiles relative to impaired-only or non-individually optimized control, also validating the inclusion of human feedback at biomechanical and algorithmic levels (Qian et al., 23 Mar 2025).
• Multi-Robot Planning and Coordination: In multi-robot coverage and temporal logic satisfaction, HITL frameworks guarantee both safety and persistent task satisfaction under human overrides, dynamic obstacles, and communication failures. Locally coordinated re-planning, MPC-based trap avoidance, and mixed-initiative controllers secure formal guarantees even in unstructured and dynamic fields (Yu et al., 2023, Shi et al., 2024).
• Air Traffic Control and Critical Infrastructure: Human-in-the-loop, regulation-aligned assessment frameworks for AI controllers in air traffic control incorporate legally-mandated scenario curation, human-instructor reporting, and formal reliability metrics, ensuring plausibility, transparency, and rigorous quantitative grounding in safety-critical environments (Carvell et al., 7 Jan 2026).
• Adaptive and Supervisory Cyber-Physical Systems: Biophysically informed models of human decision-making, strategy-selection, and stress/fatigue effects enable supervisory controllers to optimally allocate decision-making resources between humans and automation, maintaining high overall system reward rates even as human operator state stochastically evolves (Firouznia et al., 2018).

6. Limitations, Challenges, and Future Directions

Despite empirical gains, HITL control frameworks face enduring theoretical and practical challenges:

• Human Model Uncertainty: Many current HITL controllers rely on simplistic or static human models, while real-world behavior is context-dependent, non-stationary, and often poorly modeled by tractable stochastic processes. Progress in Markov chain + fuzzy-inference integration and learning from large-scale human-in-the-loop logs (such as in (Banerjee et al., 2024)) is partly alleviating these issues, but further advances require richer multi-modal and longitudinal data.
• Operator Workload and Authority Allocation: Simulator and real-world studies highlight that fixed arbitration strategies (e.g., constant α blending) are suboptimal across populations and urgency scenarios; individual differences in workload, trust, and reactivity demand adaptive, context-sensitive authority assignment and pre-operation training (Zhou et al., 2021, Mishra et al., 15 Jul 2025).
• Safety Certification with Stochastic Actions: Achieving rigorous, probabilistic forward-invariance and system safety in the presence of non-linear, multi-factor stochastic human behaviors is computationally expensive and often requires combinatorial explosion control (see the use of mixture reduction, branch-and-bound MIQP in (Choi et al., 2022, Shi et al., 2024)).
• Scalability and Real-Time Constraints: Branch-and-bound, kernel-based regression, and digital-twin simulation are effective up to moderate system sizes, but further advances are needed in distributed optimization, asynchronous human-machine protocols, and communication-efficient architectures to meet the scaling demands of field-deployed systems (Mishra et al., 15 Jul 2025, Shi et al., 2024).

HITL control continues to be a primary focus of research in robotics, robotics–human interaction, automation, and cyber-physical systems, driven by the imperative to maintain robust, adaptive, safe, and human-aligned performance in increasingly complex and uncertain application environments. The convergence of advanced learning, simulation, and human modeling methodologies is expected to expand the capabilities and reliability of such frameworks in the coming years.