Map-Indexed Performance Funnels for Robots Crossing Into Unfamiliar Regions

A robot or car tuned in one place often behaves worse somewhere else. Road friction, slope, lane width and wind differ from region to region, and a controller tuned for home conditions can overshoot or leave its lane before it adapts. Prescribed performance control (PPC) promises the opposite of "it adapts eventually." It is a family of tracking controllers that keep the tracking error inside a user-chosen envelope, called a funnel, at every instant, not just in the long run. The catch is that the funnel is a fixed function of time that only shrinks. The guarantee also quietly assumes the actuators can always push hard enough. When a vehicle enters a region whose disturbances exceed its spare actuator authority, the controller demands unbounded effort, saturates, and the guarantee fails.

We propose to make the funnel a function of where the robot is, not only of how long it has been running. A location-indexed estimate of how large the disturbances are, with uncertainty that is high in unvisited places, sets how fast the funnel is allowed to shrink. Lane geometry, extracted from map or aerial data, sets how wide the funnel is allowed to be. Where the narrowest feasible funnel is wider than the lane permits, the controller has proven ahead of time that it cannot guarantee staying in lane at the planned speed. It then slows the reference before arriving, instead of discovering the problem by saturating.

Core Claim

The claim is narrow and testable. When disturbances jump abruptly at a region boundary, for example where asphalt turns to cobblestone or ice, widening the funnel before the boundary from a map of disturbance bounds keeps the tracking error inside its envelope. A purely reactive design, which widens only after an online disturbance observer has noticed the change, does not. Neither does any fixed-time funnel. The advantage should grow with the size of the jump relative to the observer's bandwidth, and should vanish for slowly varying disturbances. If it does not vanish there, we will have measured the wrong thing.

The mechanism rests on one observation. The PPC barrier only requires the funnel $\rho(t)$ to stay positive and smooth with a bounded derivative. Monotone shrinking is a design convention, not a requirement. Widening a funnel never endangers the barrier, because the ratio of error to funnel width only falls. Shrinking is what consumes control authority.

Take error dynamics $\dot e = f + g u + d$ with input limit $|u|\le u_{\max}$, known model-error bound $|f|\le F$, and a disturbance bound $\bar D(p)$ that depends on the position $p$. The largest rate at which the error can be driven toward zero at the funnel boundary is $a(p) = g\,u_{\max} - F - \bar D(p)$. We therefore require

$$\dot\rho(t) \ge -\big(a(p(t)) - \delta\big), \qquad \rho_{\min}(p) \le \rho(t) \le w(p),$$

where $\delta>0$ is a safety margin, $\rho_{\min}(p)$ is the smallest steady-state width the residual disturbance allows, and $w(p)$ is the lane half-width minus the vehicle half-width. The disturbance bound is $\bar D(p) = \mu(p) + \beta\sigma(p)$. Here $\mu$ and $\sigma$ are the mean and standard deviation of a Gaussian process (a regression model that reports its own uncertainty) fit over position and map covariates, and $\beta$ is a confidence multiplier.

This borrows directly from Provably Robust Learning-Based Approach for High-Accuracy Tracking Control of Lagrangian Systems. That work uses Gaussian process variance to enlarge the robustness bound in unvisited states of a robot arm, but it proves only eventual boundedness, with no transient envelope. The research file flags a tension: that work's bounds must expand under uncertainty while PPC funnels must contract (Gap 2). The constraint above resolves the tension by letting expansion be unconditional and contraction be rate-limited.

Research Program

The work has three parts.

The first is analysis. We will show that a standard error transformation $e/\rho = \tanh(\epsilon)$, combined with a backstepping or sliding-mode law and a funnel obeying the constraint above, keeps the transformed error $\epsilon$ bounded even when $\bar D$ jumps discontinuously. The condition is that the jump is announced by the map at least one funnel-relaxation time ahead. We will also state plainly what happens otherwise: an unannounced jump larger than $\delta$ can still break the guarantee.

The second is the infeasibility certificate. Whenever $\rho_{\min}(p) > w(p)$ along the upcoming path, the planner's reference is retimed to a lower speed until feasibility returns. This turns a silent failure into a planning decision.

The third is a set of method choices borrowed from prior systems. The disturbance observer follows the adaptive sliding-mode observer of Disturbance Rejection Control for Autonomous Trolley Collection Robots with Prescribed Performance. That work held tracking within 0.01 m in a single simulated airport environment but assumed a fixed disturbance-derivative bound, which is exactly what fails at region boundaries (Gap 4). Our design lets the observer's estimate update the map online. For allocation under actuator limits, we follow the quadratic-program reference allocation of Prespecified-Performance Kinematic Tracking Control for Aerial Manipulation. We will also study the input-constrained PPC of Adaptive neural network control of a 2-DOF helicopter system considering input constraints and global prescribed performance, which handles saturation for a single, spatially uniform plant. We treat it as a constraining prior: our contribution is the spatial indexing and the anticipation, not saturation handling in general.

On the geographic side, cross-city planner adaptation has so far been judged open-loop. SkyDrive: Learning to Drive in a New City from Aerial Traffic Monitoring fine-tunes planners on drone-derived trajectories from Songdo. It cut the RAP planner's final displacement error from 8.116 m to 2.730 m, but never executes those waypoints with a low-level controller. Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving shows why that matters: low open-loop displacement error predicts closed-loop success poorly, and AD-MLP reaches 0% success. We use SkyDrive-style references as the thing being tracked. The question is whether a residual error of a few metres can be executed in lane by a controller that knows where tracking will be hard.

Baselines

The naive reference point is a nominal tracking controller tuned in the source region and deployed unchanged. We use pure pursuit plus a PID speed loop, tuned on source towns with the same budget as every other method. Any method that fails to beat it on funnel violations in target regions has no claim.

The strongest PPC competitor is fixed-time PPC with an adaptive sliding-mode observer, implemented as in Disturbance Rejection Control for Autonomous Trolley Collection Robots with Prescribed Performance. Its funnel parameters will be tuned on the source regions with the same tuning budget we give our method. Its reported number is a tracking bound within 0.01 m on a simulated circular trajectory. It reports no cross-region result, and the file has no reported result on Lacuna for fixed-funnel PPC under cross-region transfer. Beating it means fewer funnel violations and less saturation time on held-out regions at matched mean tracking error.

The strongest learning-based robust controller is the Gaussian-process robust controller of Provably Robust Learning-Based Approach for High-Accuracy Tracking Control of Lagrangian Systems. It reports a 41.5% RMS error reduction over nominal control on a UR10 arm. We give it the same Gaussian process over position and the same source data. It isolates the question of whether a funnel adds anything beyond variance-scaled robustness. Beating it means a lower worst-case lateral error and fewer lane departures, not merely lower RMS.

The strongest constraint-native competitor is adaptive nonlinear model predictive control (MPC), a controller that re-solves a short-horizon constrained optimization at every step. It gets lane constraints and the same disturbance map. In DATT: Deep Adaptive Trajectory Tracking for Quadrotor Control, adaptive MPC tracked an infeasible star trajectory with 0.11 m error against DATT's 0.05 m, and needed more than four times DATT's sub-3.2 ms inference time. We run MPC at the same control rate as our method and, separately, at its native rate. Beating it means comparable violation rates at a fraction of the compute, or fewer violations at equal compute.

Three ablations isolate the key ingredient. Reactive widening drives the funnel constraint from the online observer alone, with no map; this is the central comparison. Boundary reset widens the funnel by a hand-chosen amount at every region boundary. Conservative constant funnel fixes the funnel at the widest value used anywhere.

Why not just…?

Why not just use MPC with lane and input constraints, refit per region? MPC handles constraints natively and is the most serious threat. Its weaknesses are compute, and hard-constraint infeasibility when a disturbance jump pushes the state outside the feasible set. The DATT latency and infeasible-trajectory numbers above suggest the first matters on embedded hardware. Our experiments run MPC with exactly the same map, so if it matches our violation rate at the same control rate, the funnel adds nothing and we will report that. Our expected edge is that the funnel gives a closed-form law with a cheap certificate rather than a per-step solve. We do not claim MPC cannot use a map.

Why not just widen the funnel reactively when the observer detects trouble? This alternative could plausibly match us, and we have built the proposal around testing it. For slow disturbance changes it should perform as well. The map earns its place only when the disturbance jump outpaces the observer. We sweep jump magnitude and observer bandwidth, and predict the curves cross. If reactive widening matches anticipatory widening everywhere, the geographic component is unnecessary and the contribution shrinks to rate-limited non-monotone funnels.

Why not just raise gains or schedule them by GPS region? High gains narrow steady error, but they saturate actuators and add chatter. Decentralized Adaptive Control with Improved Transient Performance pursues transient improvement through gain design without hard envelopes. The trolley study found sliding-mode control produced harsh control signals where PPC stayed smooth. Gain scheduling is close in spirit to our boundary-reset ablation, which tests it directly.

Why not just make the planner region-invariant, as SkyDrive or auxiliary-depth training does? Better representations reduce error but do not bound it. In DoShiCo: a Domain Shift Challenge for Control, auxiliary depth lifted real-world maneuver accuracy only from 35% to 64%. SkyDrive's adapted planner still has a 5.89% time-to-collision infraction rate. These approaches are complementary: we track their references.

Proposed Experiments

The main testbed is CARLA through Bench2Drive. Its 12 towns serve as regions: we tune on a subset and deploy on held-out towns. Within the held-out towns we assign per-region tire friction and grade patches with abrupt boundaries, across Bench2Drive's 23 weather conditions. The map is built from a small number of prior passes, and its uncertainty is genuinely high off those passes.

The primary metrics are three. Funnel-violation rate is the fraction of runs in which error leaves $\rho$. Lane-departure rate is measured against $w(p)$. Saturation duty cycle is the fraction of time at $u_{\max}$. We also report RMS lateral error, worst-case lateral error, and the Bench2Drive driving score. For the latter the strongest reported reference is DriveAdapter's 42.08% mean ability score; no reported result found on Lacuna for funnel-violation rate on Bench2Drive.

A second experiment tracks SongdoDrive-derived references from SkyDrive-adapted planners. We replay the drone-extracted lane geometry as $w(p)$ and run a vehicle dynamics model in closed loop. The question is whether the reported 2.730 m final displacement error can be executed in lane, and how often the certificate correctly predicts the places where it cannot.

The decisive check is the jump-magnitude sweep against reactive widening. Map-indexed anticipation should cut violations sharply for jumps larger than the margin $\delta$ arriving faster than the observer's time constant, and tie elsewhere.

A second decisive check is calibration. Among segments the certificate flags as infeasible, unconstrained execution should violate the lane at a high rate. Among segments it passes, violations should stay near zero. A certificate that flags everything would make the method merely conservative, so we report false-alarm rate alongside missed violations.

Finally, we deliberately corrupt the map, both by shifting friction patches and by understating $\bar D$. This measures how gracefully the guarantee degrades when the geographic prior is wrong. Mis-specified priors are the realistic failure mode in a new city, and the method is only useful if it fails no worse than the reactive baseline when its map is wrong.