When designing controllers for autonomous systems—whether a self-driving car navigating a track or a robotic arm assembling components—we often rely on finite-horizon optimization. At each time step, the system looks ahead a few seconds into the future, plans an optimal sequence of actions, executes the very first action, and then repeats the process at the next step.
This "greedy" look-ahead strategy works remarkably well in practice, but it harbors a dangerous trap: the short-horizon myopia. Imagine driving a vehicle toward a sharp cliff. If your planning horizon is only two seconds ahead, and it takes three seconds of hard braking to come to a complete stop, a short-horizon controller might decide to accelerate for the next half-second because the vehicle is "perfectly safe" within its immediate two-second window. By the time the next planning cycle occurs, the vehicle has entered an unrecoverable state where a future crash is mathematically inevitable.
As noted by Alexander Liniger et al. in Real-Time Control for Autonomous Racing Based on Viability Theory, local optimization strategies often fail to account for long-term safety, leading systems into states that are safe in the short term but make future failures unavoidable. To prevent this, control theorists rely on a property called recursive feasibility (or persistent feasibility). Recursive feasibility is the mathematical guarantee that if a valid, constraint-satisfying control plan exists at the current time step, a valid plan is guaranteed to exist for all subsequent time steps.
The Terminal Constraint Solution
We cannot solve the short-horizon trap by simply planning infinitely far into the future; the computational cost of infinite-horizon optimization is prohibitive for real-time systems. Instead, recursive feasibility allows us to achieve infinite-horizon safety guarantees within a finite-horizon optimization framework.
The core mechanism relies on introducing a terminal constraint at the end of the finite planning window. Instead of letting the final state of the look-ahead horizon float freely, we force it to land within a pre-verified controlled invariant set—a "safe zone" from which we know a valid control policy can keep the system safe indefinitely.
Current State ----> [ Planned Trajectory ] ----> Terminal State ∈ Controlled Invariant Set
│
▼
(Infinite-Horizon Safe Policy)
By ensuring the local plan terminates in this safe zone, we guarantee that the system always has a viable fallback option. Hao Wang et al. demonstrate this principle in Safe and Performant Deployment of Autonomous Systems via Model Predictive Control and Hamilton-Jacobi Reachability Analysis by using a safety value function derived from Hamilton-Jacobi reachability analysis as a terminal constraint. This ensures that even if a high-level planner (such as a hallucinating Large Language Model) provides suboptimal or erroneous instructions, the low-level predictive controller can always find a safe path forward.
Similarly, Alexander Liniger et al. demonstrate in Safe Motion Planning for Autonomous Driving using an Adversarial Road Model that incorporating a viability-based safe set as a terminal constraint allows for significantly shorter prediction horizons. The terminal state acts as a proxy for the infinite horizon, keeping the optimization computationally lightweight yet provably safe.
Maintaining Feasibility Under Uncertainty and Learning
Proving recursive feasibility becomes substantially more complex when a system must learn online or operate in uncertain environments. If an adaptive controller updates its internal model parameters to improve performance, it risks moving the system into a state where no future safe path exists under the updated model (Lukas Hewing et al. in Learning-Based Model Predictive Control: Toward Safe Learning in Control).
To preserve recursive feasibility amidst noise and model shifts, researchers introduce structural constraints that bound the system's uncertainty:
- Constraint Tightening: In data-driven control where an explicit mathematical model is unavailable, noise can easily cause future constraint violations. Julian Berberich et al. prove in Robust Constraint Satisfaction in Data-Driven MPC that systematically tightening the optimization constraints based on noise bounds guarantees that a valid solution will persist across all subsequent steps.
- Monotonic Refinement: When updating probability distributions online, Peter Coppens demonstrates in Data-Driven Distributionally Robust MPC for Constrained Stochastic Systems that constraining new uncertainty ambiguity sets to be subsets of prior sets ensures the controller remains safe under all previously considered distributions.
- Monotonicity Constraints on Tubes: When deep learning is used to predict safety boundaries ("tubes") around a robot's path, Ali Agha shows in Deep Learning Tubes for Tube MPC that enforcing a monotonicity constraint on predicted uncertainty bounds prevents the safety tube from shrinking arbitrarily. This ensures that a valid control solution at one timestep mathematically implies the existence of a solution at the next.
Anytime Feasibility and Data-Driven Safe Zones
In high-speed applications like autonomous racing, a real-world controller might experience transient computational bottlenecks, causing the optimization solver to fail to return a new solution before the next control cycle.
To achieve robust safety under computational delays, Sean Vaskov et al. design a framework in Guaranteed Safe Reachability-based Trajectory Design for a High-Fidelity Model of an Autonomous Passenger Vehicle where every planned trajectory is constrained to terminate in a guaranteed safe state, such as a full stop. If the real-time planner fails to find a new valid path within the allocated time, the system safely executes the pre-verified braking maneuver from the previous planning cycle.
Alternatively, Logan Numerow et al. introduce an "anytime feasible" approach in Inherently robust suboptimal MPC for autonomous racing with anytime feasible SQP. By constraining the planning horizon to terminate on a precomputed, feasible periodic reference trajectory, the controller always plans toward a known-good state. This ensures that the system remains within safe operating bounds for multiple control cycles even if the solver fails to produce an update.
For repetitive or iterative tasks, these safe zones can be constructed directly from experience rather than complex analytical geometry.
Iteration 1: [ Successful Trajectory A ] ──┐
├──> Convex Hull = Data-Driven Terminal Set
Iteration 2: [ Successful Trajectory B ] ──┘
As shown by Hannes Petrenz et al. in Robust MPC for Uncertain Linear Systems - Combining Model Adaptation and Iterative Learning and Lukas Brunke in RLO-MPC: Robust Learning-Based Output Feedback MPC for Improving the Performance of Uncertain Systems in Iterative Tasks, taking a convex combination of states from past successful iterations creates a valid terminal constraint. By treating past safe trajectories as reachable targets, the controller ensures recursive feasibility and stability, allowing the system to safely push its performance limits over time.