Compound Feedforward-Feedback Control

Updated 28 May 2026

Compound feedforward-feedback control is a two-degree-of-freedom architecture that integrates model inversion feedforward with corrective feedback to achieve improved tracking and robustness.
The feedforward component anticipates control actions using models, lookup tables, or learning algorithms, while the feedback loop compensates for any residual errors and disturbances.
Empirical validations in aerospace, robotics, and distributed systems demonstrate that combining both components outperforms using either feedforward or feedback alone in terms of transient response and stability.

Compound feedforward-feedback control refers to a two-degree-of-freedom (2DOF) control architecture in which feedforward and feedback controllers operate in parallel (or, more rarely, in series), jointly shaping the system’s response. The feedforward component typically leverages a model of the plant to invert or anticipate desired outputs, directly compensating for nominal system dynamics, references, or known disturbances. The feedback loop acts on the output or state error, correcting for unmodeled dynamics, disturbances, uncertainties, nonlinearity, or sensor noise, and enforcing stability, robustness, and constraint satisfaction. Compound feedforward-feedback structures are widely employed in tracking problems for nonlinear underactuated systems, robotic and mechatronic platforms, safety-critical aerospace applications, cyber-physical systems (e.g., traffic and resource networks), and advanced learning-augmented control. Empirical evidence establishes that the combination outperforms either component alone, producing enhanced transient tracking, disturbance rejection, and robustness to model errors and sampling limitations (Drücker et al., 2023, Sayyed et al., 19 Jan 2026, Raach et al., 2021, Pedroso et al., 2023, Brummelhuis et al., 2020, Zhang et al., 2022).

1. Formal Structure and Principle of Superposition

In classical form, the compound 2DOF control law for a SISO or MIMO system can be written as

$u(t) = u_{ff}(t) + u_{fb}(t),$

where $u_{ff}$ is the feedforward term and $u_{fb}$ is the feedback term. The feedforward block maps the reference $r(t)$ (and possibly disturbances or measured exogenous signals) through an inverse or predictive model of the plant, yielding a control input that (ideally) brings the system close to the target trajectory in the absence of disturbances and uncertainties. The feedback block corrects any residual error between the system output and target, compensating for model mismatch, unmodeled plant phenomena, or unexpected disturbances.

A generic compound architecture appears in numerous domains:

System class	Feedforward $u_{ff}$	Feedback $u_{fb}$
Mechanical tracking	Servo-inverse, DAE, or lookup table	Nonlinear/linear feedback
Reset/precision motion control	Adaptive inverse	Reset element (CI, SORE)
Learning-based/SL policy	RL agent, policy network	SMC, LQR, PI, etc.
Distributed resource allocation	KKT derivative/predictor	Fixed-time/high-gain law

Closed-loop signals remain bounded as long as the feedback gains are tuned to dominate the residual uncertainty after feedforward compensation (Drücker et al., 2023, Sayyed et al., 19 Jan 2026).

2. Feedforward Design: Model Inversion, Prediction, and Adaptation

Feedforward design exploits models that encode known plant physics, reference trajectories, or disturbance profiles:

Model Inversion and Servo Constraints: For mechanical systems, feedforward may solve a differential-algebraic system that includes reference-tracking as explicit algebraic constraints. For minimum-phase plants, implicit integration (e.g., Euler) yields a real-time $u_{ff}$ sequence that achieves output matching modulo model errors (Drücker et al., 2023).
Lookup Table or Explicit Map: Static or dynamically parameterized actuator settings are tabulated over operating points, e.g., engine speed and fuel injection for combustion engine airpath control (Zhang et al., 2022).
Learning-Augmented Feedforward: In nonlinear and underactuated settings, RL agents—typically parameterized as deep neural networks—are trained to approximate the optimal feedforward policy using reward shaping and safety filters that enforce input constraints (Sayyed et al., 19 Jan 2026).
Adaptive or Online Identification: In precision motion platforms and reset control, feedforward parameters are recursively adapted via gradient-like schemes, leveraging error signals and specialized detection of convergence intervals to avoid learning during unstable transients (Brummelhuis et al., 2020).
Distributed Prediction: In large-scale or networked systems, distributed feedforward laws integrate consensus-based estimators to locally reconstruct the time-derivative of the KKT optimality conditions, generating per-agent feedforward terms that anticipate how the optimizer drifts in time (Xu et al., 2024).

Nominal feedforward achieves exact tracking only if the plant is perfectly modeled and all states and disturbances are known. Any model mismatch produces residual error, which propagates to feedback (Drücker et al., 2023, Zhang et al., 2022).

3. Feedback Design: Robustness, Disturbance Rejection, Stability

Feedback controllers in compound architectures are responsible for stabilizing the closed-loop system, constraining the tracking error, and absorbing model uncertainties that the feedforward cannot handle:

Funnel/Barrier Feedback: For underactuated or light-weight mechanical systems, a funnel controller enforces the error to remain within time-varying, pre-defined bounds (the "funnel"), with a high-gain law that raises the control authority as the error approaches the boundary (Drücker et al., 2023).
Robust/Sliding-Mode Feedback: In aerospace applications, sliding-mode control (SMC) law robustly enforces stability under hard input constraints and modeling errors, with Lyapunov-certifiable tuning. An SMC layer can be overlaid atop a learning-augmented policy to “salvage” performance of partially trained policies (Sayyed et al., 19 Jan 2026).
Classical PI/PID and Shaped Feedback: For systems with strong nonlinearities or non-smooth hysteresis (e.g., magnetic shape memory (MSM) actuators), a PI controller with explicit loop-shaping is employed alongside feedforward inverse hysteresis compensation, closely tracking the desired output while attenuating sensing noise via filtering (Ruderman et al., 25 Feb 2025).
Distributed/Decentralized Correction: In time-varying resource allocation, per-agent and dual-variable feedback laws drive the stationarity residuals to zero in finite time, guaranteeing global feasibility and constraint satisfaction (Xu et al., 2024).
Reset Control: Reset feedback elements (Clegg integrator, SORE) introduce nonlinear phase compensation or rapid error correction, especially in repetitive or high-precision motion (Brummelhuis et al., 2020).

Performance of the closed-loop system depends critically on gain scheduling, boundary-setting (as in funnel or SMC), and the allocation of corrective authority between feedback and feedforward. Robustness to actuator lag, noise, and sample-rate reductions is maintained as long as feedback “dominates” the model uncertainty/residual disturbance (Drücker et al., 2023, Sayyed et al., 19 Jan 2026).

4. Experimental Realizations and Case Studies

Empirical validation across multiple domains underscores the efficiency and flexibility of compound feedforward-feedback schemes:

Underactuated Mechanical Systems: On a two-flywheel torsional oscillator, combined funnel-feedforward control reduced transient tracking error (as measured by $\int e(t)^2$ over $[0,10]$ s) and steady error energy (over $[10,15]$ s) beyond what either pure feedforward or pure funnel control could achieve; chattering/spiky control signals were substantially mitigated, and stability at reduced sample rates was preserved (Drücker et al., 2023).
Aerospace Control (6DOF Flight): RL+SMC hybrid controllers for F-18 models delivered near-RL transient speed (spin recovery <6 s), but strictly bounded overshoot ( $u_{ff}$ 0 rad), chattering suppression ( $u_{ff}$ 1 reduction), and zero rate/deflection limit violations even with partially trained policies—the SMC feedback “safety net” being essential for robustness (Sayyed et al., 19 Jan 2026).
Wind Farm Coordination: Combining FLORIS-based surrogate-model FF yaw angles with local PI feedback using lidar-based wake detection yielded a $u_{ff}$ 2 total energy gain (vs baseline) and robust adaptation to changing inflow or model errors (Raach et al., 2021).
Motion Control with Hysteresis: MSM actuator setups using inversion-free KP-based feedforward and classic PI feedback improved RMS and peak tracking error by $u_{ff}$ 3 compared to either component alone, especially under high sensor noise (Ruderman et al., 25 Feb 2025).
Urban Traffic Signal Networks: In large road networks, FF (from exogenous demand estimation) plus FB (from link occupancy) LQ controllers reduced total time spent (TTS) and relative queue imbalances by up to $u_{ff}$ 4– $u_{ff}$ 5 under demand surges, outperforming either component alone (Pedroso et al., 2023).

Tabulated results from select experimental validations:

Study/Domain	Metric	FF Only	FB Only	Compound FF+FB
Mechanical oscillator (Drücker et al., 2023)	$u_{ff}$ 6 (trans)	high	moderate	lowest
Wind farm (Raach et al., 2021)	ΔEnergy (%)	11.2	—	16.9
MSM actuator (Ruderman et al., 25 Feb 2025)	RMS error (1Hz)	0.050mm	0.030mm	0.009mm
Traffic network (Pedroso et al., 2023)	TTS (w/ surge)	—	—	–10–15% vs. FB-only
Aerospace (Sayyed et al., 19 Jan 2026)	Chattering ( $u_{ff}$ 7)	high	moderate	lowest

5. Theoretical Guarantees and Design Trade-Offs

Compound feedforward-feedback control architectures admit a range of formal performance and stability guarantees, subject to the plant assumptions and control law selection:

Closed-Loop Boundedness & Stability: Superposition of exact model inversion (feedforward) and high-gain or barrier function (feedback) yields well-posedness—provided the feedback gain overtakes worst-case residual error after feedforward (Drücker et al., 2023).
Robustness to Input Constraints and Disturbances: Sliding-mode and Lyapunov-based feedback layers can formally guarantee sliding condition and finite-time convergence even with bounded feedforward mismatch and actuator rate/position limits (Sayyed et al., 19 Jan 2026, Xu et al., 2024).
Explicit Disturbance/Uncertainty Handling: In state-space linear systems, LQG separation theorem and its feedforward extensions establish that the optimal policy decomposes into a sum of feedback on the Kalman state estimate and feedforward on exogenous/reference/disturbance estimators (Zhang et al., 2023).
Finite-Time and Fixed-Time Convergence: Distributed resource allocation and high-order sliding-mode feedback enable convergence to optimal trajectories within a predetermined, uniform time bound, independently of initial errors (Xu et al., 2024).

Design trade-offs are inherent:

Large feedforward authority can reduce error but must be counteracted by proportionally higher feedback gain or restricted via safety filters to maintain Lyapunov stability.
Tight funnel/boundary settings (small c in $u_{ff}$ 8) sharpen steady-state tracking at the expense of higher transient feedback gain and potential sensitivity to sampling artifacts (Drücker et al., 2023).
Online feedforward computation (e.g., real-time DAE solving) imposes computational overhead, mitigated by efficient integrators and model reduction, but may limit applicability in low-latency or resource-constrained scenarios.

6. Advanced Architectures and Learning-Augmented Extensions

Emerging research has expanded the scope and structure of compound feedforward-feedback control:

Learning-Based Hierarchical Control: Policy-structure neural networks separate feedback and feedforward roles, enabling structured learning and transfer between modules, reducing RL sample-complexity (Zhang et al., 2023, Kobayashi et al., 2021). Mixture policies can adaptively weight feedback and feedforward, ensuring robust operation under partial observability or sensor failure (Kobayashi et al., 2021).
Distributed and Projection-Based Algorithms: Decentralized resource allocation schemes combine distributed KKT-based feedforward (consensus-predicted optimizer drift) with fixed-time sliding-mode–style feedback, using projection to handle per-agent state constraints; the architecture generalizes to cyber-physical networks with evolving feasibility sets (Xu et al., 2024).
Multi-Loop Visual-Inertial Fusion: In sensor-fusion for pose estimation, coordinated gradient-descent feedback (e.g., Madgwick’s correction), feedforward initialization (roll-pitch from IMU), and bias correction feedback loops jointly stabilize and robustify orientation, outperforming filter-only or direct-smooth approaches (Chen et al., 2020).

A key insight from these advanced frameworks is that enforcing structural separation (e.g., LQG-type theorems) or modular network design can yield both interpretability and accelerated convergence in high-dimensional, nonlinear, or partially observable domains (Zhang et al., 2023, Kobayashi et al., 2021).

7. Limitations, Current Challenges, and Future Directions

Combining feedforward with feedback does not eliminate all control challenges:

Model Dependence: Compound structures depend on the availability and accuracy of plant models for feedforward synthesis; significant unmodeled dynamics or rapid time-varying phenomena may erode gains unless feedback gain is increased, which can interact adversely with sample rate and noise (Drücker et al., 2023, Zhang et al., 2022).
Online Computational Load: Real-time inversion (as in servo-constraint or DAE-based feedforward) or adaptive schemes may impose prohibitive computational requirements, particularly at high bandwidth. Model reduction, sparse sampling, and parallelization are active mitigation strategies (Drücker et al., 2023, Brummelhuis et al., 2020).
Tuning Interactions: Overlap in authority can induce conflicts (e.g., high feedback gain destabilizing if feedforward “overshoots”), requiring careful tuning, authority blending (as in SMC “zeta” scheduling), and explicit safety filtering (Sayyed et al., 19 Jan 2026).
Generalization and Learning: While learning-based controllers have demonstrated successful integration of feedforward policy learning and robust feedback overlays, generalization beyond the training domain and formal certification under non-stationarity or sensor failures remain open research issues (Zhang et al., 2023, Kobayashi et al., 2021).

Future research is focused on sampled-data/funnel-MPC variants, direct learning of feedforward models in real time, blended model-predictive/feedback-funnel designs, and applications to high-dimensional MIMO systems with distributed architectures (Drücker et al., 2023, Xu et al., 2024, Zhang et al., 2022). Advances in embedding formal convergence and robustness properties into neural-network–based control structures remain an area of active exploration.

Key References

Funnel+feedforward for underactuated and sampled-data tracking (Drücker et al., 2023)
RL-synthesized feedforward with SMC/Lyapunov feedback in flight control (Sayyed et al., 19 Jan 2026)
Distributed feedback-feedforward resource allocation algorithms (Xu et al., 2024)
Adaptive feedforward for reset feedback motion control (Brummelhuis et al., 2020)
Compound FF-FB wake redirection in wind farm optimization (Raach et al., 2021)
Model predictive airpath control with FF in diesel engines (Zhang et al., 2022)
RL-structural separation of feedforward-feedback in observer-controller (Zhang et al., 2023)
Direct acoustic impedance regulation by mixed FF-FB (Volery et al., 2021)
Compound control in sensor-fusion and pose estimation (Chen et al., 2020)
Trajectory-level stochastic optimal FF-FB with uncertainty integration (Berret et al., 5 Mar 2026)
Feedback-feedforward optimization in RL for robustness to sensing loss (Kobayashi et al., 2021)