Newton-Raphson Flow Tracking Controller

• Newton-Raphson flow-based tracking controllers are prediction-based variable-gain controllers that recast output tracking as an online root-finding problem.
• The controller evolves the input through a continuous-time Newton correction embedded in an auxiliary differential equation to match predicted future outputs.
• Extensive validation in simulations and hardware, including mobile robotics and autonomous vehicles, highlights its efficacy and design tradeoffs.

Searching arXiv for recent and foundational papers on Newton-Raphson flow-based tracking control.
A Newton-Raphson flow-based tracking controller is a prediction-based variable-gain integral controller that recasts output tracking as an online root-finding problem for a future-output matching equation. For a nonlinear plant
[
\dot{x}(t)=f(x(t),u(t)), \qquad y(t)=h(x(t)),
]
the controller predicts a future output (\hat y(t+T)=g(x(t),u(t))) under a held input, compares it with the future reference (r(t+T)), and evolves the control input through a continuous-time Newton correction rather than a discrete Newton iteration. In this sense, the controller is “flow-based”: the control input is itself the state of an auxiliary differential equation implementing a Newton step in real time [1708.04117][1910.00693]. Later work specialized the construction to differentially flat systems, where the predictive map becomes explicit in flat coordinates and the tracking law can be written in closed form [2508.12694][2605.29231].

1. Origins and conceptual identity

The modern formulation emerged from lookahead-simulation tracking, where a nonlinear memoryless Newton-Raphson idea was extended to dynamic plants by replacing the unavailable static map (y=g(u)) with a predicted future output map computed from the plant model [1708.04117]. In the memoryless case, the controller solves (g(u)=r); in the dynamic case, it instead solves
[
g(x(t),u(t))=r(t+T),
]
where (g) denotes the output predicted (T) seconds ahead while the current input is held constant. This shift from instantaneous matching to predictive matching is the defining structural move of the method [1811.08033].

Subsequent work made the Newton-flow interpretation explicit. The control law was written as a continuous-time variable-gain integrator whose gain is the inverse sensitivity of predicted future output with respect to the current input, together with a controller speedup parameter (\alpha) that sharpens convergence and, in some cases, stabilizes the closed loop [1910.00693]. In that formulation, the controller is not an algebraic inversion law and not a receding-horizon optimizer; it is a dynamical system for (u(t)) driven by a predictive residual.

A further line of development addressed systems with differential flatness. Rather than designing a standard flatness-based feedforward tracker, these works used the flat representation primarily to build a predictor and a tractable Newton update. The resulting controller is still Newton-Raphson in the sense of predictor inversion or Newton correction, but its tracking guarantees are formulated in flat coordinates and then related back to the original nonlinear dynamics [2508.12694][2605.29231].

2. Predictive root-finding formulation

The core predictive equation is
[
\hat y(t+T)=g(x(t),u(t)),
]
with (T>0) a lookahead horizon. In the general nonlinear setting, one constructs (g) by introducing a predictor state (\xi(\tau)) on ([t,t+T]) satisfying
[
\dot{\xi}(\tau)=f(\xi(\tau),u(t)), \qquad \xi(t)=x(t),
]
and then setting
[
\hat y(t+T)=h(\xi(t+T)).
]
Tracking is therefore posed as the online root-finding problem
[
r(t+T)-g(x(t),u(t))=0
]
rather than as a direct instantaneous equality (r(t)-y(t)=0) [1910.00693].

With this predictive map, the basic Newton-Raphson flow controller is
[
\dot{u}(t)=\alpha \left(\frac{\partial g}{\partial u}(x(t),u(t))\right)^{-1}\bigl(r(t+T)-g(x(t),u(t))\bigr),
]
assuming the input and output have the same dimension and the Jacobian (\frac{\partial g}{\partial u}) is nonsingular along the closed-loop trajectory [1811.08033][1910.00693]. Equivalent sign conventions also appear, for example
[
\dot{u}(t)= - \alpha\bigg(\frac{\partial g}{\partial u}(x(t), u(t))\bigg)^{-1}(g(x(t), u(t)) - r(t+T)),
]
which is the same correction law written in terms of the residual (g-r) rather than (r-g) [2508.12694].

A more exact predictive formulation augments the basic Newton step with chain-rule compensation terms:
[
\dot{u}(t)=\left(\frac{\partial g}{\partial u}(x(t),u(t))\right)^{-1} \Bigl( r(t+T)-\hat y(t+T) +\dot r(t+T) -\frac{\partial g}{\partial x}(x(t),u(t))f(x(t),u(t)) \Bigr).
]
For the predicted residual
[
e_p(t)=r(t+T)-g(x(t),u(t)),
]
this yields the exact first-order error dynamics
[
\dot e_p(t)=-e_p(t),
]
and, with
[
V(x(t),u(t))=\frac12 |r(t+T)-\hat y(t+T)|^2,
]
the Lyapunov identity
[
\dot V(x(t),u(t))=-2V(x(t),u(t)).
]
This is one of the clearest formulations of the Newton-flow idea: the inverse predictor Jacobian maps a future-output residual into a differential input correction, and the compensation terms linearize the predicted residual dynamics exactly [1910.00693].

This structure explains two persistent features of the method. First, the controller primarily regulates a predicted future output, not the instantaneous plant output. Second, its convergence is not the quadratic convergence theory of discrete Newton iteration; under exact compensation it is a continuous-time residual-damping flow with first-order exponential error dynamics [1910.00693].

3. Differential-flat realization

For a class of differentially flat systems, the plant output is chosen as a subset of the state,
[
y:= [x_{p_1}, x_{p_2}, \hdots, x_{p_M}]^\top,
]
and the flat output is assumed to coincide with the plant output:
[
y_f(t)=y(t)=h(x(t)).
]
The nonlinear system is related by invertible (C^1) transformations to a flat linear system
[
\dot z = Az + Bv,\qquad y_f = Cz,
]
with
[
z=\Phi(x), \qquad v=\Gamma(x,u), \qquad x=\phi(z), \qquad u=\gamma(z,v).
]
In this setting, the flat model is used primarily as a predictor rather than as a conventional feedforward flatness controller [2508.12694].

Under the same held-input idea, the flat predictor becomes
[
\tilde z(t+T)=e^{AT}z(t)+\int_0^T e^{A(t-\tau)}d\tau \, Bv(t):=Rz(t)+Sv(t),
]
so that
[
g_f(z(t),v(t)):=C\tilde z(t+T)=CRz(t)+CSv(t).
]
The Jacobian is simply
[
\frac{\partial g_f}{\partial v}=CS,
]
which makes the Newton correction and, in the static limit, direct predictor inversion explicit [2508.12694].

This leads to two closely related controllers. The first is the Dynamical Newton-Raphson Controller (DNRC),
[
\dot{u}(t) = - \alpha\bigg(\frac{\partial g}{\partial u}(x(t), u(t))\bigg)^{-1}(g(x(t), u(t)) - r(t+T)),
]
which is the genuine flow-based controller. The second is the Statical Newton-Raphson Controller (SNRC) obtained in the singular limit (\alpha=\infty), where the predictor equation is enforced instantaneously:
[
g(x(t),u(t))=r(t+T).
]
For the flat predictor this yields
[
v(t)=(CS)^{-1}(r(t+T)-CRz(t)),
]
and hence
[
u(t)=\gamma!\left(\Phi(x(t)),(CS)^{-1}(r(t+T)-CR\Phi(x(t)))\right).
]
The tracking theory in the flatness paper is developed for this static infinite-speed limit rather than for the finite-(\alpha) DNRC [2508.12694].

A later flatness paper made the same viewpoint explicit for vehicle models. The flat-output dynamics were written as chains of integrators, the Newton correction was applied to the flat input (\nu), and the physical controls were recovered through the inverse endogenous map. Under additional assumptions on the flatness transformation, the (\alpha)-dependent terms coincide with those obtained by applying the Newton-Raphson controller directly to the nonlinear dynamics, while any discrepancy is isolated in an explicit drift term of the form (\frac{\partial u}{\partial \tilde y}\dot{\tilde y}) [2605.29231].

4. Stability, convergence, and error characterization

The most complete closed-form convergence results occur in the memoryless setting. For
[
y(t)=g(u(t)),
]
and the flow
[
\dot{u}(t)=\alpha \left(\frac{\partial g}{\partial u}(u(t))\right)^{-1}\bigl(r(t)-g(u(t))\bigr),
]
if
[
\eta:=\sup_{t\ge 0}|\dot r(t)|,
]
then
[
\limsup_{t\to\infty}|r(t)-g(u(t))|\le \frac{\eta}{\alpha},
]
and for constant references one obtains asymptotic regulation. The same (1/\alpha) structure reappears in the predictive setting, but only after separating prediction error from controller-induced residual error [1708.04117][2508.14185].

For the predictive nonlinear controller with exact chain-rule compensation, the predicted residual obeys
[
\dot e_p(t)=-e_p(t),
]
and with additive compensation error (\mathcal E_2(t)) one gets
[
\limsup_{t\to\infty}|r(t)-\hat y(t)|\le \frac{\eta_2}{\alpha},
]
where (\eta_2:=\limsup_{t\to\infty}|\mathcal E_2(t)|). If prediction error is
[
\mathcal E_1(t)=\hat y(t+T)-y(t+T), \qquad \eta_1:=\limsup_{t\to\infty}|\mathcal E_1(t)|,
]
then the actual output tracking error satisfies
[
\limsup_{t\to\infty}|r(t)-y(t)|\le \eta_1+\frac{\eta_2}{\alpha}.
]
A common misunderstanding is that increasing (\alpha) eliminates all asymptotic error; it does not. It attenuates the compensation or reference-variation term, but not the irreducible prediction mismatch (\eta_1) [1910.00693].

The flatness-based stability results are more specialized. For regulation with finite (\alpha), local asymptotic stability is proved in a neighborhood of the origin if the corresponding flat closed-loop and predictor satisfy a verifiable stability criterion, and the paper gives the semi-quantitative estimate
[
\delta < K_S \leq \alpha K_L
]
for the radius of the local domain of attraction [2508.12694]. For tracking, however, the main theorem concerns only the static infinite-speed limit. Under a uniqueness assumption for predictor inversion, bounded references, and the Hurwitz condition
[
A - B(CS)^{-1}CR \ \text{is Hurwitz},
]
the SNRC achieves tracking in terms of prediction:
[
\hat y(t+T)=r(t+T).
]
This is not a direct theorem that (y(t)\to r(t)); it is a theorem that the predicted output matches the future reference exactly, while the flat closed loop remains bounded [2508.12694].

The same paper quantifies the gap between predictive tracking and actual tracking by introducing
[
e_p(x,u):=g(x,u;t,T)-h(x), \qquad e_f(x,u):=g_f(\Phi(x),\Gamma(x,u);t,T)-g(x,u;t,T),
]
and proving that for sufficiently small (T),
[
|e_p(x,u)|\le K_PT,\qquad |e_f(x,u)|\le K_FT.
]
Thus both predictor mismatch and flat-predictor mismatch are (O(T)). This formalizes the practical rule that smaller lookahead horizons make predictive matching more representative of true output tracking, although excessively small (T) may compromise stability in finite-(\alpha) closed loops [2508.12694].

A later flatness paper addressed the (\alpha)-stability notion more directly. For trivial flat-output dynamics, it derived sufficient conditions under which the flat closed loop is (\alpha)-stable, in particular when the flat-output dynamics form chains of integrators of sufficiently low order, and then demonstrated the resulting controllers on the kinematic unicycle and dynamic bicycle models [2605.29231].

5. Implementations and application domains

The method has been implemented in simulation and hardware on mobile robots, cars, quadrotors, and a blimp. In autonomous-vehicle studies, the controller was applied to a nonlinear dynamic bicycle model by numerically integrating both the predictor state and the sensitivity equation for (\frac{\partial g}{\partial u}). Reported simulation results include closed-curve tracking with lateral error peaks of (2\,\text{cm}), (5\,\text{cm}), and (8\,\text{cm}) at (15\,\text{km/h}), (25\,\text{km/h}), and (35\,\text{km/h}), respectively, and lane-change tracking with lateral peaks of (7\,\text{cm}), (16\,\text{cm}), and (25\,\text{cm}) at (10\,\text{m/s}), (15\,\text{m/s}), and (19\,\text{m/s}). In a Robotarium platoon experiment, after (t=1.4\,\text{s}), the average control error over four robots was approximately (5\,\text{mm}) [1811.08033].

In intersection traffic control, the same predictive Newton-flow tracker was layered beneath a higher-level trajectory planner and then combined with control barrier functions. In a tracking-only experiment with a deliberate (100\%) mass mismatch in the predictor, the reported tracking errors were below about (6\,\text{cm}) during the initial transient and then below (2\,\text{cm}) after roughly (3) seconds. When barrier functions were activated, the controller preserved a (5\,\text{m}) minimum inter-vehicle distance and reduced maximum lateral deviation to about (0.27\,\text{m}), at the cost of sacrificing perfect tracking when safety constraints became active [2004.10226].

Aggressive quadrotor tracking gave a hardware demonstration of the flow-based architecture with a simple hover-linearized predictor. The controller used
[
\dot u(t) = \alpha \left( \frac{\partial p}{\partial u}(x(t),u(t)) \right)^{-1} \bigl(r(t+T)-p(x(t),u(t))\bigr),
]
with output (y=[p_x,p_y,p_z,\psi]^\top), and on a Holybro x500v2 quadrotor it outperformed the native PX4 cascaded controller on all reported benchmark trajectories. The reported RMS errors were (0.051) versus (0.26644) on a vertical circle, (0.16834) versus (0.61125) on a horizontal circle, and (0.10508) versus (0.27228) on a fast vertical tall lemniscate. The ROS2 control node published at (100\,\text{Hz}), and the average computation time for the Newton-Raphson flow controller with I-CBF constraints was (7.138\times 10^{-5} \pm 2.971\times 10^{-5}\,\text{s}) [2408.11197].

A later aerial study emphasized computational constraints. On a miniature blimp, the Newton-Raphson flow controller achieved the best reported tracking RMSE on every tested trajectory, for example (0.07866) versus (0.14075) for NMPC and (0.19332) for feedback linearization on Circle A, with compute times around (0.84)–(0.96\,\text{ms}) at a required (40\,\text{Hz}) update rate. On a quadrotor, the same controller was generally within the (10\,\text{ms}) budget required by a (100\,\text{Hz}) offboard loop, while the acados-based NMPC often consumed almost the entire budget or exceeded it. The paper also reported that the Newton-Raphson controller had the lowest CPU energy expenditure for both the blimp and quadrotor platforms [2508.14185].

The flatness-based papers use the same architecture on lower-dimensional vehicle models but with explicit predictor inversion. For the kinematic bicycle tracking example in the stability paper, the reference
[
r(t)=\begin{bmatrix}20\sin(0.05t)\30\sin(0.1t)\end{bmatrix}
]
was tracked with three horizons (T=0.5\,\text{s}), (5.0\,\text{s}), and (10.0\,\text{s}). As (T) increased, tracking error increased and input transients decreased; with (T=0.5), the reported tracking error (|y(t)-r(t)|) was less than (3\,\text{cm}) [2508.12694].

6. Limitations, design tradeoffs, and related directions

The central design tradeoff is between prediction horizon (T) and controller speedup (\alpha). Smaller (T) makes the predictor more faithful to the true near-future output, as reflected by the (O(T)) bounds on predictor mismatch, and often improves tracking fidelity. At the same time, several papers note that too small a lookahead horizon can destabilize the closed loop for a fixed (\alpha), whereas increasing (\alpha) can restore stability or reduce the asymptotic residual [1708.04117][2508.14185]. The tuning problem is therefore not a conventional fixed-gain problem: (T) changes the predictive map itself, and (\alpha) changes the time scale of the Newton flow.

A second limitation is the structural requirement that the Jacobian of predicted output with respect to input be invertible. The main theory uses the actual inverse, not a pseudoinverse. This entails square input-output dimensions, nonsingularity of (\frac{\partial g}{\partial u}), and potentially severe operating-region restrictions. In the flatness-based bicycle example, the inverse input transformation is invertible iff (V\neq 0), so the practical validity region excludes stopping. In the quadrotor and blimp studies, no separate Jacobian regularization or Levenberg-Marquardt damping was introduced; practical constraint handling was instead delegated to barrier-function modifications of the nominal flow [1910.00693][2408.11197].

A third limitation concerns the scope of existing proofs. The most rigorous nonlinear predictive tracking result with exact residual dynamics depends on chain-rule compensation terms [1910.00693]. The flatness-based stability paper proves local regulation for the finite-(\alpha) dynamical controller but analyzes tracking only for the static infinite-speed limit. It therefore does not provide a full nonlinear finite-(\alpha) tracking theorem for the flow controller on that class of systems [2508.12694]. Likewise, the later flatness paper proves (\alpha)-stability for the trivial flat dynamics under sufficient conditions and demonstrates encouraging vehicle simulations, but it explicitly notes that convergence of the nominal controller is experimentally observed rather than fully proved [2605.29231].

The controller also differs in important ways from neighboring paradigms. It is not MPC, because it does not solve an online optimization over a control sequence; it predicts one future output under a frozen input and applies one Newton-style correction. It is not feedback linearization, because it does not require a full normal-form transformation or exact cancellation of plant nonlinearities. It is not ordinary integral control, because the integrator gain is the inverse local sensitivity of a predicted-output map. These distinctions are explicit in both the theoretical and experimental literature [1811.08033][2408.11197].

A related but distinct direction is discrete-time online Newton tracking. The Online Sketched Newton-Raphson method tracks time-varying roots or optimizers with one sketched Newton correction per sample and extends naturally to equality-constrained problems such as optimal power flow. That framework is not a continuous-time Newton flow, but it can be read as a discrete-time analogue of Newton-based tracking around moving operating points, with computational savings obtained by sketching and tracking guarantees expressed through dynamic regret and constraint-violation bounds [2605.27985].

In mature form, the Newton-Raphson flow-based tracking controller is therefore best understood as a family of predictive input-dynamics laws organized around a single principle: solve, in real time, the equation “predicted future output equals future reference,” and do so by embedding Newton’s inverse-Jacobian correction inside the feedback loop. The main theoretical questions that remain open are the full nonlinear finite-(\alpha) tracking theory, systematic handling of singular predictor Jacobians, and robustness guarantees that explicitly quantify the effect of model mismatch, numerical prediction error, and actuator constraints.