Real-time control of autonomous systems requires solving complex optimization problems within milliseconds. While physical systems are inherently nonlinear and non-convex, our most efficient numerical solvers rely on linear models. To bridge this gap, engineers frequently use local linearization—approximating complex dynamics or safety boundaries using first-order Taylor expansions around a nominal operating point.
However, as Julian Berberich et al. note in Linear tracking MPC for nonlinear systems Part II: The data-driven case, a system's true behavior inevitably deviates from these linear predictions as it moves away from its current operating point. This mismatch introduces linearization errors that severely compromise both safety and path optimality.
The Cost of Linearization: Infeasibility vs. Conservatism
When a control framework simplifies nonlinear constraints into linear ones, it introduces a fundamental trade-off. If the approximation is too optimistic, it creates a safety gap; if it is too pessimistic, it induces crippling conservatism.
The Safety Gap (Under-conservatism)
If a linear approximation forms an outer approximation of the safe space, the planner may choose a trajectory that is mathematically valid according to the model but physically dangerous. For instance, Luigi Palmieri et al. demonstrate in An NMPC Approach using Convex Inner Approximations for Online Motion Planning with Guaranteed Collision Avoidance that linearizing obstacle constraints can cause a robot to mathematically appear safe while physically colliding with an obstacle.
Similarly, during aggressive maneuvers, unmodeled nonlinearities can cause a system to drift away from its planned safe trajectory (Kai Zhong et al. in Fast Second-order Cone Programming for Safe Mission Planning), potentially leading to catastrophic constraint violations in physical infrastructure like power grids (Spencer Hutchinson et al. in Safe Pricing Mechanisms for Distributed Resource Allocation with Bandit Feedback).
The Conservatism Trap (Over-conservatism)
To guarantee safety despite these errors, traditional frameworks often enforce strict, static safety buffers or linear half-spaces. However, as Shuo Liu et al. observe in Iterative Convex Optimization for Model Predictive Control with Discrete-Time High-Order Control Barrier Functions, approximating nonlinear safe sets with linear planes artificially excludes large, valid portions of the safe state space.
This reduction of the feasible search area leads to highly conservative trajectories, preventing autonomous agents from navigating narrow passages or dense obstacle fields (Vivek K. Adajania et al. in AMSwarmX: Safe Swarm Coordination in CompleX Environments via Implicit Non-Convex Decomposition of the Obstacle-Free Space). In extreme cases, these large error bounds cause the safety filter to unnecessarily override task actions, resulting in the "frozen robot" problem where a valid path theoretically exists but cannot be found (Kai-Chieh Hsu et al. in ISAACS: Iterative Soft Adversarial Actor-Critic for Safety; Akash Ratheesh in Operator Splitting Covariance Steering for Safe Stochastic Nonlinear Control).
The Core Mechanism: Bounding the Taylor Remainder
The core insight of this research direction is that linearization error should not be treated as unmodeled noise or managed via ad-hoc safety margins. Instead, it must be treated as a deterministic, bounded disturbance derived directly from the system's underlying mathematics.
Any sufficiently smooth nonlinear function $f(x)$ can be expressed exactly using a first-order Taylor expansion plus a Lagrange remainder term:
$$f(x) = f(x_0) + \nabla f(x_0)(x - x_0) + e(x)$$
Here, $e(x)$ represents the exact linearization error. As Arshiya Taj Abdul et al. demonstrate in Nonlinear Robust Optimization for Planning and Control, by treating this remainder $e(x)$ as a bounded uncertainty, the optimization framework can explicitly account for the inaccuracies introduced during successive linearization.
This approach fundamentally changes how safety margins behave:
- State-Dependent Margins: Because the remainder $e(x)$ depends on the distance from the linearization point $(x - x_0)$, the safety margins are dynamically scaled. When the system executes mild maneuvers near $x_0$, the error bound shrinks, allowing the controller to utilize more of the available state space.
- Guaranteed Robustness: By designing a controller that satisfies constraints for the worst-case value of $e(x)$, the system achieves formal safety guarantees even under highly nonlinear edge cases where standard linear approximations fail (Thom Badings et al. in Probabilities Are Not Enough: Formal Controller Synthesis for Stochastic Dynamical Models with Epistemic Uncertainty; Hai Zhong et al. in Chance-Constrained Iterative Linear-Quadratic Stochastic Games).
Integrating Error Bounds into Robust Control
To translate these mathematical error bounds into physical trajectories, researchers embed the bounded remainders into robust control and verification pipelines:
- System Level Synthesis (SLS): Traditional robust control can be overly cautious when handling parametric uncertainties and linearization errors simultaneously. To combat this, Antoine P. Leeman et al. propose a framework in Robust Optimal Control for Nonlinear Systems with Parametric Uncertainties via System Level Synthesis that jointly optimizes the vehicle's nominal trajectory and its feedback controller. This co-design significantly reduces conservatism while maintaining robust stability.
- Reachability Analysis: In complex systems like power grids, manual linear approximations can compromise verification precision (Matthias Althoff et al. in ARCH-COMP18 Category Report: Continuous and Hybrid Systems with Linear Continuous Dynamics). By combining reachability analysis with robust optimization, Yichen Zhang et al. show in Robust trajectory-constrained frequency control for microgrids considering model linearization error how to guarantee frequency stability by explicitly accounting for the mathematical gap between the simplified model and physical reality.
- Hardware Safety Buffers: When applying linearized safety filters to real-world hardware, unmodeled remainder terms can still cause minor boundary violations. Trager Joswig-Jones demonstrates in Safe Control of Grid-Interfacing Inverters with Current Magnitude Limits that introducing explicit, mathematically derived safety buffers into linearized Control Barrier Functions prevents hardware damage without requiring computationally prohibitive fully-nonlinear control formulations.
Key Foundational Work
Fast Second-order Cone Programming for Safe Mission Planning (2016) — contributed 1 concept elements to this direction.
Region of attraction estimation for a perching aircraft: A Lyapunov method exploiting barrier certificates (2012) — contributed 1 concept elements to this direction.
Controller synthesis for linear dynamical systems with adversaries (2016) — contributed 1 concept elements to this direction.