Reactive Model Predictive Contouring Control

• Reactive Model Predictive Contouring Control (RMPCC) is a path-following framework that formulates motion generation as a receding-horizon optimal control problem parameterized by path progress.
• It decomposes tracking error into contouring and lag components, enabling the controller to dynamically trade off accuracy, progress, and safety via embedded barrier functions.
• Real-time performance is achieved using Jacobian linearization and Gauss-Newton approximations, as demonstrated in 7-DoF manipulators and other MPCC applications.

Searching arXiv for the cited RMPCC and related MPCC papers to ground the article in current literature. Reactive Model Predictive Contouring Control (RMPCC) is a path-following framework that formulates motion generation as a receding-horizon optimal control problem over a reference path parameter rather than a fixed time parametrization. In this formulation, the controller optimizes both the system motion and the evolution of path progress online, allowing the controlled system to trade off contouring accuracy, lag along the path, and progress in the presence of constraints, disturbances, and dynamic hazards. In the robot-manipulation literature, RMPCC has been presented as an NMPC-based framework that successfully avoids obstacles, singularities and self-collisions in dynamic environments at 100 Hz, using Control Barrier Functions (CBFs), Jacobian-based linearization, and a Gauss-Newton Hessian approximation (Yoon et al., 13 Aug 2025). Closely related MPCC formulations in quadrotors, vehicles, bipedal robots, excavators, industrial machines, and articulated commercial vehicles establish the broader conceptual basis of reactive contouring control: path-parametric prediction, online time allocation or progress selection, and explicit handling of safety and feasibility constraints (Romero et al., 2021, Krinner et al., 2024, Bertipaglia et al., 2024, Bertipaglia et al., 2024, Narkhede et al., 2023, Sotiropoulos et al., 2021, Yuan et al., 2024, Yuan et al., 13 Apr 2026, Aertssen et al., 27 Apr 2026).

1. Conceptual definition and scope

RMPCC belongs to the MPCC family, but emphasizes reactive path following in dynamic, constraint-rich environments. The defining distinction is that the reference is parameterized by a scalar path parameter rather than time. In the manipulator formulation, the reference path is written as $p_r(s)$ in $\mathbb{R}^3$, with $s \in [0,1]$ denoting the path parameter and $v_s$ its velocity; the optimization therefore decides not only the robot motion but also how fast the motion should progress along the path (Yoon et al., 13 Aug 2025). The same structural idea appears in quadrotor MPCC, where arc length parameter $\theta$ is augmented as an additional state, and progress along the reference path is directly optimized online rather than prescribed by a pre-computed schedule (Romero et al., 2021).

This path-parametric view is central to reactivity. Conventional time-based parameterization imposes a strict time schedule for path progression and can lead to degraded tracking or abrupt motions when obstacle avoidance, singularity avoidance, or other constraints force deviations. In the manipulator case, the path-parameterized formulation allows the system to slow, pause, or even reverse $s$ when dynamic obstacles or singularities are encountered (Yoon et al., 13 Aug 2025). In the bipedal Digit application, the same principle allows the robot to decide between faithful versus fast path following and to overtake a moving obstacle by temporarily deviating from the path when contouring error is weighted less heavily than lag error (Narkhede et al., 2023). This suggests that “reactive” in RMPCC denotes not merely online replanning, but online reshaping of the spatiotemporal realization of path traversal.

A second defining feature is the explicit decomposition of tracking error into contouring and lag components. In MPCC literature, contouring error quantifies orthogonal deviation from the path, while lag error measures deviation along the tangent direction of the path. This decomposition is used in quadrotor flight (Romero et al., 2021), bipedal locomotion (Narkhede et al., 2023), excavator control (Sotiropoulos et al., 2021), articulated vehicles (Aertssen et al., 27 Apr 2026), and robot manipulators (Yoon et al., 13 Aug 2025). RMPCC inherits this structure and combines it with safety constraints that operate directly on the predicted state trajectory.

2. Core mathematical structure

In the manipulator formulation, the state and control inputs are

$$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$

where $q \in \mathbb{R}^n$ are the robot joint angles, $\dot{q}$ are the joint velocities, $s$ is the path parameter, and $\mathbb{R}^3$0 is its velocity. The dynamics are written in discretized form as

$\mathbb{R}^3$1

using a linear time-invariant model as derived in the source formulation (Yoon et al., 13 Aug 2025).

For a reference path $\mathbb{R}^3$2 and end-effector position $\mathbb{R}^3$3, the contouring and lag errors are defined by projection onto the path tangent $\mathbb{R}^3$4: $\mathbb{R}^3$5

$\mathbb{R}^3$6

The orientation error is expressed through the $\mathbb{R}^3$7 logarithmic map,

$\mathbb{R}^3$8

The paper notes that an algebraic approximation is used for the contouring term in practice (Yoon et al., 13 Aug 2025).

The finite-horizon cost is

$\mathbb{R}^3$9

This objective balances path-following accuracy, path-progress regulation, orientation tracking, joint-velocity regularization, rate smoothness, and path-parameter acceleration (Yoon et al., 13 Aug 2025).

The corresponding optimal control problem is posed as

$s \in [0,1]$0

The same general structure recurs across MPCC variants: augmented progress dynamics, contouring and lag error penalties, and a progress-promoting term. In quadrotor MPCC, for example, the stage cost contains contouring error, lag error, regularization terms, and a term $s \in [0,1]$1 that maximizes progress along the path (Romero et al., 2021). In excavator MPCC, the augmented progress state is $s \in [0,1]$2 with update $s \in [0,1]$3, and the cost includes contouring and lag errors, a progress reward $s \in [0,1]$4, and regularization on inputs and progress changes (Sotiropoulos et al., 2021). This consistency indicates that RMPCC is best understood as a reactive specialization within the broader path-parametric MPCC class.

3. Safety constraints and barrier-based reactivity

A principal characteristic of RMPCC is the use of Control Barrier Functions to encode safety-relevant geometric constraints directly in the prediction problem. In the manipulator framework, CBFs are used for singularity avoidance, self-collision avoidance, and obstacle or environment avoidance (Yoon et al., 13 Aug 2025). For a differentiable function $s \in [0,1]$5, the safe set is

$s \in [0,1]$6

and the CBF condition is

$s \in [0,1]$7

where $s \in [0,1]$8 is class $s \in [0,1]$9 (Yoon et al., 13 Aug 2025).

The manipulator paper adopts a relaxed form,

$v_s$0

with

$v_s$1

where $v_s$2 is a quadratic function for regularization near zero (Yoon et al., 13 Aug 2025). This allows barrier constraints to remain numerically tractable inside the NMPC loop.

The specific barrier functions are: $v_s$3 for singularity avoidance,

$v_s$4

for self-collision avoidance, with $v_s$5 predicted by a neural network over joint configurations, and

$v_s$6

for obstacle or environment avoidance, where $v_s$7 is obtained from a Neural Signed Distance Function (Neural-JSDF) (Yoon et al., 13 Aug 2025).

The reactive effect of these constraints derives from the fact that the controller may adjust $v_s$8, $v_s$9, and $\theta$0 online rather than being forced to satisfy a fixed timing law. The manipulator experiments show modulation of $\theta$1, $\theta$2, and $\theta$3 for different constraint-driven avoidance events (Yoon et al., 13 Aug 2025). A parallel mechanism appears in the bipedal MPCC with moving-obstacle avoidance, where a barrier condition

$\theta$4

is applied to the obstacle-distance function

$\theta$5

to ensure safe separation during overtaking (Narkhede et al., 2023).

Related MPCC formulations encode safety in other ways. In quadrotor racing, MPCC++ introduces a prismatic tunnel-shaped spatial constraint and a terminal set so that safety is enforced by track constraints while time optimization remains in the cost (Krinner et al., 2024). In aggressive vehicle obstacle avoidance, the controller uses emergency-weighted obstacle-distance penalties, road-edge terms, and friction-circle constraints, so obstacle avoidance “takes over” as the vehicle nears an obstacle (Bertipaglia et al., 2024). These variants are not labeled RMPCC in the same way as the manipulator work, but they illustrate the same design principle: reactivity arises when path progress is optimized jointly with explicit spatial safety constraints.

4. Numerical solution and real-time implementation

RMPCC is computationally demanding because it combines nonlinear kinematics, geometric path-following terms, orientation tracking on $\theta$6, and barrier constraints. The manipulator framework attains 100 Hz by combining Jacobian-based linearization with a Gauss-Newton Hessian approximation and sequential quadratic programming (Yoon et al., 13 Aug 2025).

The cost terms, including contouring, lag, and orientation errors, are linearized via first-order Taylor expansions. For contouring and lag error, the Jacobian has the structure

$\theta$7

while the orientation-error Jacobian uses an approximation of the $\theta$8 log map with the right Jacobian $\theta$9 (Yoon et al., 13 Aug 2025). The Gauss-Newton approximation forms the Hessian as $s$0 for the least-squares objective terms, avoiding the expense of exact second derivatives. The nonlinear OCP is then solved via SQP, with OSQP used as the QP solver. Table III reports mean computation time per control step at 8.71 ms, which is sufficient for 100 Hz (Yoon et al., 13 Aug 2025).

Comparable real-time strategies appear in other MPCC systems. Quadrotor MPCC is implemented using ACADO with a real-time iteration scheme and QPOASES, with demonstrated solve times of approximately 5 ms per iteration for horizon length $s$1, enabling 100 Hz feedback (Romero et al., 2021). The bipedal Digit implementation solves its MPCC at 15 Hz with sub-15 ms solve times (Narkhede et al., 2023). Vehicle MPCC for obstacle avoidance uses FORCESPro’s nonlinear interior-point solver with a 30-step prediction horizon and 0.05 s sample time (Bertipaglia et al., 2024). Articulated commercial-vehicle MPCC is solved with IPOPT and CasADi in receding horizon form, with approximately 100 ms per iteration for the reported settings (Aertssen et al., 27 Apr 2026). These implementation choices vary by application, but all reflect the need to make path-parametric optimal control sufficiently fast for closed-loop replanning.

A plausible implication is that RMPCC’s practical viability depends less on a single solver architecture than on the ability to expose a structured least-squares geometry, exploit differentiable models, and approximate nonconvexities in a way that preserves sufficiently accurate prediction without destroying the solve budget.

5. Reported performance in robot manipulation

The manipulator RMPCC framework is evaluated on a 7-DoF Franka Panda in real and simulated environments while following a 3D lemniscate path and avoiding singularities, self-collision, and a moving spherical obstacle (Yoon et al., 13 Aug 2025). The reported metrics are contouring error $s$2, orientation error $s$3, end-effector acceleration $s$4, and the safety quantities $s$5, $s$6, and $s$7. The time plots show that manipulability, self-distance, and environment distance remain above safety thresholds throughout execution (Yoon et al., 13 Aug 2025).

The experiments compare RMPCC against TT-MPC. Across all scenarios, RMPCC maintains contouring errors below $s$8 maximum and $s$9 mean, whereas TT-MPC peaks at $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$0 in the reported comparisons (Yoon et al., 13 Aug 2025). Orientation errors under RMPCC are described as two orders of magnitude lower. The framework also keeps end-effector acceleration low, reported as below $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$1, versus above $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$2 with TT-MPC (Yoon et al., 13 Aug 2025). The control loop runs stably at 100 Hz with 8.7 ms average per step, and the approach is reported to outperform state-of-the-art methods by a factor of 10 (Yoon et al., 13 Aug 2025).

The following reported scenario-level results summarize the comparison:

Scenario	Controller	Selected reported result
Self-collision	RMPCC	Max acceleration $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$3, max $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$4, max $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$5
Singularity	RMPCC	Max acceleration $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$6, max $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$7, max $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$8
Environment collision	RMPCC	Max acceleration $$x = \begin{bmatrix} q \ s \ v_s \end{bmatrix} \in \mathbb{R}^{n+2}, \quad u = \begin{bmatrix} \dot{q} \ \dot{v}_s \end{bmatrix} \in \mathbb{R}^{n+1},$$9, max $q \in \mathbb{R}^n$0, max $q \in \mathbb{R}^n$1

These results are specific to the reported setup, but they clarify the central empirical claim: RMPCC couples low contouring error with reactive safety handling and low acceleration, rather than treating evasive action as an external supervisory process (Yoon et al., 13 Aug 2025).

6. Relation to adjacent MPCC variants

RMPCC is most directly represented in the manipulator literature, yet its constituent ideas have appeared across several MPCC lines of work.

In quadrotor flight, MPCC concurrently solves time allocation and control. The controller augments the state with path progress $q \in \mathbb{R}^n$2, minimizes contouring and lag errors, and maximizes progress via $q \in \mathbb{R}^n$3, so the optimal times at which points on the path are reached are chosen online (Romero et al., 2021). MPCC++ extends this formulation with a track constraint, terminal set, residual dynamics learned from real-world data, and Trust Region Bayesian Optimization (TuRBO) for hyperparameter tuning. It achieves similar lap times to the best-performing RL policy, outperforms the best model-based controller while satisfying constraints, and consistently prevents gate crashes with 100% success rate in simulation and real world (Krinner et al., 2024). Although the quadrotor papers do not foreground the term RMPCC, they exemplify reactive contouring control in the sense of online progress allocation and hard spatial safety.

In automated driving, several MPCC-based obstacle-avoidance controllers expose another branch of reactive contouring control. One nonlinear MPCC for vehicle obstacle avoidance combines motion planning, path tracking and vehicle stability objectives, prioritising collision avoidance in emergencies, and uses torque vectoring to generate an extra yaw moment during evasive maneuvers at the limit of handling (Bertipaglia et al., 2024). A related learning-based MPCC introduces an online Student-t Process to capture mismatches between the prediction model and measured lateral tyre forces and yaw rate, feeding posterior means into the prediction model and propagating posterior covariances into lateral velocity and yaw rate over the horizon (Bertipaglia et al., 2024). In high-fidelity simulation, that controller avoids obstacles, keeps the vehicle stable during a double lane change at $q \in \mathbb{R}^n$4, succeeds at an 8.5% higher speed than a classical MPCC, and reduces the peak sideslip angle by 76% compared to the GP-based design (Bertipaglia et al., 2024). These formulations show that reactive contouring control can also be uncertainty-aware or stability-aware rather than only geometry-aware.

In bipedal robotics, MPCC for Digit chooses footsteps that maximize path traversal while permitting online decisions about speed and path fidelity. When contouring error is weighted less than lag error, the robot can suggest an off-path detour to overtake a moving obstacle and return afterward, whereas Cartesian tracking behavior remains blocked behind the obstacle (Narkhede et al., 2023). In excavation, a Koopman-DFL lifting linearization yields a lifted linear dynamic model that enables MPCC to be posed as a convex QP while remaining reactive to soil-profile changes and interaction forces (Sotiropoulos et al., 2021). In articulated commercial vehicles, MPCC is extended to multiple anchor points with scenario-dependent weighting and explicit corridor constraints for the front and rear tractor axles and the semitrailer axle, allowing both forward and reverse motion and highlighting the importance of semitrailer constraints for jackknifing avoidance (Aertssen et al., 27 Apr 2026). These applications indicate that RMPCC is not tied to a single embodiment but to a control architecture.

7. Guarantees, interpretations, and open distinctions

A recurring misconception is to treat contouring control as equivalent to ordinary trajectory tracking. The literature distinguishes them sharply. Time-parametrized tracking fixes when the system should be where on the path, whereas contouring control makes path progress itself a decision variable. In the manipulator work, this is explicitly presented as crucial when the error between the desired path and actual position becomes large during evasive maneuvers (Yoon et al., 13 Aug 2025). In quadrotors, the same idea is cast as solving time allocation and control concurrently (Romero et al., 2021). In biped locomotion, it determines whether the robot trails an obstacle or overtakes it (Narkhede et al., 2023).

Another distinction concerns “minimizing” contouring error versus “guaranteeing” a contouring bound. Standard MPCC and RMPCC formulations generally optimize contouring and lag errors subject to constraints, but some industrial contouring-control works make bounded contouring error the central objective. For biaxial switched linear systems, a contouring error-bounded control algorithm uses MPC together with switch control-invariant sets to guarantee state, input, and contouring error constraints for any admissible mode switching, with recursive feasibility and closed-loop stability (Yuan et al., 2024). For biaxial systems with structural flexibility and input delay, robust control invariant sets and delay-augmented models are used so that the true contouring error $q \in \mathbb{R}^n$5 satisfies

$q \in \mathbb{R}^n$6

thereby delivering a path-agnostic bound on contouring error (Yuan et al., 13 Apr 2026). These results are not presented as RMPCC, but they mark an adjacent line in which hard contouring guarantees take precedence over online progress optimization.

The present literature therefore supports a narrow and a broad reading of RMPCC. In the narrow reading, RMPCC denotes the manipulator framework that couples path-parameterized NMPC with CBFs, Jacobian-based linearization, and Gauss-Newton approximations for 100 Hz reactive path following (Yoon et al., 13 Aug 2025). In the broader reading, it denotes MPCC formulations whose path progress is optimized online and whose safety handling is embedded inside the predictive controller rather than delegated to a separate planner or mode switch. The collected evidence across manipulators, quadrotors, vehicles, bipeds, excavators, and industrial biaxial machines suggests that the broad reading is technically coherent, but only the manipulator paper uses the term as the explicit framework name (Yoon et al., 13 Aug 2025).