Feedback Guidance in Control and Robotics

• Feedback Guidance (FBG) is a control strategy that employs online sensor and learning feedback to dynamically adjust system performance across diverse applications.
• It integrates measurement-based cues and predictive signals to improve accuracy, safety, and adaptability in fields such as robotics, medical intervention, and diffusion-based generative models.
• FBG systems combine multimodal feedback with adaptive control laws, enabling enhanced interception, landing, and human-computer interaction under uncertain and dynamic conditions.

Feedback Guidance (FBG) is a broad class of control and information-delivery strategies in which system feedback—obtained via sensors, evaluators, or learning modules—is used to provide real-time or context-specific corrective signals, cues, or controls that guide system evolution or human actions toward a desired objective. In technical domains, FBG manifests in diverse forms, including robotic surgical navigation, adaptive control of complex physical systems, diffusion model guidance in generative AI, sensor network protocols, and educational toolkits. The principal feature unifying FBG methods is the use of online, measurement-based or predictive feedback to dynamically adjust guidance, as opposed to static, pre-computed, or purely feedforward protocols.

1. Feedback Guidance in Robotic Systems and Medical Intervention

FBG strategies in robotic systems tightly couple perception and action, leveraging real-time feedback from specialized sensors to enhance task accuracy, safety, and operator efficacy. A canonical application is collaborative robotic biopsy, where dynamic kinesthetic and trajectory feedback assist the user in precise needle placement. In "Collaborative Robotic Biopsy with Trajectory Guidance and Needle Tip Force Feedback" (Mieling et al., 2023), the system architecture integrates preoperative imaging, 7-DOF robotic alignment, and insertion constrained along a planned axis. Critical to the approach is an optical coherence tomography (OCT)-based "smart needle," with a cGRU–CNN real-time inference module yielding an estimate of axial tip force $F_r(t)$ at 200 Hz. This tip force is rendered directly to the operator via an admittance-style haptic interface, forming a closed kinesthetic feedback loop: $$x_d(t) = K_i \int_0^t e_F(\tau) d\tau + K_p e_F(t), \qquad e_F(t)=F_H(t) - \alpha F_r(t)$$ where $F_H$ is operator handle force and $\alpha$ is an amplification gain.

This design enables the detection of tissue transitions with high sensitivity—91% event detection in user studies—and subcentimeter localization error, independent of frictional shaft forces or prior model knowledge. Limitations include the need for improved friction compensation, lateral-force sensing (future FBG or combined OCT+FBG modalities), and robustness extensions for soft-tissue targets beyond ex vivo phantoms.

2. Feedback Guidance in Adaptive and Learning-based Control

FBG underpins data-driven, model-free shape and motion control in soft and continuum robotics. Using multi-core Fiber Bragg Gratings (FBGs) for dense, distributed strain sensing, the system in "FBG-Based Online Learning and 3-D Shape Control of Unmodeled Continuum and Soft Robots in Unstructured Environments" (Lu et al., 2022) continuously reconstructs full-3D robot shape at 25 Hz and closes the loop with an online, composite adaptive controller. The unknown kinematic map between actuation commands $q$ and sensed shape features $x$ is approximated via RBF neural networks; adaptation laws guarantee global convergence in the presence of unknown external disturbances: $$\dot{\widehat{\overline W}} = \Gamma_W^{-1}\,\mathrm{proj}\Bigl\{ k_e\,Q(q̇)\,\Theta(q)\,e + k_x\,Q(q̇)\,\Theta(q)\,\tilde x + k_r\,Q(q̇)\,\Theta(q)\,r_x \Bigr\}$$ The method achieves zero asymptotic tracking error under payload, collision, and compliance variation, substantially outperforming classical constant-curvature or fixed-Jacobian approaches.

FBG's extension to tactile skill transfer is illustrated by TacCap (Xing et al., 3 Mar 2025), a wearable FBG-thimble for robust, high-bandwidth contact and force measurement, enabling successful transfer of human grasp demonstrations to robots with minimal tactile data mismatch.

3. FBG in Feedback-based Guidance Laws for Interception and Landing

FBG in guidance and control is instantiated in feedback linearization and advanced robust control architectures for interception and terminal landing. In "Feedback Linearization-based Guidance Law for Guaranteed Interception" (Dorsey et al., 9 Sep 2025), feedback linearization transforms nonlinear pursuer-evader kinematics into a linearized error system for the controlled output (e.g., range $r$ or LOS rate $\dot{\theta}$). The linearized input $v=-k_R r$ leads to a nonlinear control law,

$$u_{\rm FBL}(x,w) = \frac{-\,\alpha(x,w)\,-\,k_R\,r}{\sin(\theta-\gamma_P)}$$

where singularities at $\sin(\theta-\gamma_P)=0$ are managed via fuzzy blending with proportional navigation. A corrected LOS-based variant uses sign correction to ensure interception in all geometries, validated in Monte Carlo simulations.

Missile and pursuit-evasion scenarios also utilize measurement feedback for robust control. In the H-infinity disturbance attenuation framework (Or et al., 2020), measurement feedback and filter and control Riccati equations are coupled to yield a closed-loop law,

$$u^*(t)=−B^\mathrm{T} X(t)[I−\gamma^{−2} Y(t)X(t)]^{−1} \hat x(t)$$

with tuning governed by disturbance attenuation ratio $\gamma$ and explicit consideration of noise, initial uncertainty, and trajectory shaping penalties.

Planetary landing guidance leverages feedback adaptation via machine learning. "Adaptive Generalized ZEM/ZEV Feedback Guidance" (Furfaro et al., 2020) integrates an actor–critic reinforcement learning loop to learn state-dependent guidance gains and time-of-flight parameters for the generalized ZEM/ZEV law, enabling feasible, real-time optimal landing subject to fuel and path constraints.

4. Feedback Guidance in Diffusion and Generative Models

FBG provides a principled, closed-loop approach to conditional generation in diffusion models, replacing static classifier-free guidance (CFG) scales with dynamically adapted, state-dependent guidance:

• In "Feedback Guidance of Diffusion Models" (Koulischer et al., 6 Jun 2025), guidance is modulated at each denoising step by estimating the informativeness of the conditional signal via a recursive posterior update, producing an adaptive $\lambda_t(x_t)$ scale: $$\mu_{\rm FBG}(x_t|c) = \mu_\theta(x_t) + \lambda_t\bigl[\mu_\theta(x_t|c) - \mu_\theta(x_t)\bigr]$$ Demonstrated on ImageNet and COCO, FBG yields substantial FID/recall improvements and adaptively increases guidance for more difficult prompts.
• "Dynamic Classifier-Free Diffusion Guidance via Online Feedback" (Papalampidi et al., 19 Sep 2025) generalizes the paradigm, using online evaluators (e.g., CLIP, discriminator, or task-specific reward networks) in latent space to select the optimal guidance schedule per step/sample via a greedy search. Human preference win rates surpass 53% for overall alignment and over 55% in capability-specific tasks, demonstrating prompt- and capability-adaptive behavior across both small-scale and SoTA diffusion models.

A theoretical unification with RLHF, test-time scaling, and reward-directed diffusion is provided in (Jiao et al., 4 Sep 2025), highlighting the equivalence of FBG (exponential tilting) to KL-regularized reward maximization and presenting an efficient RL-free resampling algorithm for both autoregressive and diffusion systems.

5. Feedback Guidance in Distributed Sensing and XR/Motion Training

In networked estimation, FBG architectures improve efficiency by enabling sensor nodes to adapt their communication protocols according to central feedback. "Advantages of Feedback in Distributed Data-Gathering for Accurate and Power-Efficient State-Estimation" (Choe et al., 16 Jul 2025) proves that feeding back the fusion-center's state estimate enables each sensor to compute its own expected information gain and schedule transmission accordingly, typically reducing total transmissions by a factor of $N$ while maintaining—or exceeding—the accuracy of conventional round-robin protocols, provided concrete timing and information-rate criteria are met.

XR-based motion guidance systems formalize FBG as corrective feedback, distinguished from motion feedforward. Yu et al. (Yu et al., 14 Feb 2024) analyze the space of visual (and potentially multimodal) FBG along dimensions of modality, temporality (real-time vs. post-hoc), content (detection, magnitude, rectification), spatial mapping, and abstraction level. The interaction between feedforward and feedback modalities is shown to crucially structure user learning, attention, and cognitive load.

6. Multimodal Feedback Guidance in Human–Computer/Robotic Interfaces

The integration of visual, haptic, and auditory feedback mechanisms enables robust guidance in task-critical settings where single-modality cues may be insufficient or unreliable. In "Multimodal Feedback for Task Guidance in Augmented Reality" (Guo et al., 2 Oct 2025), a system combining optical see-through AR with wrist-based vibrotactile cues demonstrates enhanced spatial accuracy (31% reduction in depth error vs. AR-only) and usability in target-touching tasks. User studies identify design principles such as spatial mapping of cues, pull-style low-duty-cycle vibration patterns, multimodal redundancy, and modularity for real-world constraints (occlusion, lighting).

Complementary approaches in procedural software engineering training employ automated testers and coverage analyzers to provide structured, incremental feedback as part of a test-driven development workflow. The MASS toolkit (Dick et al., 30 Nov 2024) exemplifies how fine-grained, coverage-based feedback, encoded in instructor-specified JSON rules, incrementally guides students toward thorough testing and correct software implementation, yielding significant reductions in error rates and improvements in code quality.

7. Limitations, Challenges, and Future Directions

Across domains, FBG systems face limitations in model generalization, evaluation reliability, sensing modality coverage, and computational or latency overhead. Many practical systems lack full vectorial force measurement (notably axial-only in OCT-needles (Mieling et al., 2023)), depend on accurate reward/evaluator function learning (Papalampidi et al., 19 Sep 2025), or are limited by the fidelity and robustness of real-time estimation (e.g., FBG strain decoding in deformable robots (Lu et al., 2022)). In complex interaction settings, ill-designed feedback can induce cognitive overload or suboptimal attentional allocation (Yu et al., 14 Feb 2024).

Open directions include multi-modal, adaptive feedback mechanisms; integration of more expressive evaluators and controllers; robust learning under uncertainty; and development of standardized, cross-domain metrics for FBG efficacy. In collaborative and assistive robotics, enhanced tactile and haptic channels, machine-learned event detection, and deeper user-in-the-loop optimization remain key challenges for next-generation feedback-guidance systems.