1National University of Sciences and Technology, Islamabad · 2New Jersey Institute of Technology
A transverse feedback linearizing controller is undefined on a singular set, such as zero speed for a unicycle or zero thrust for a quadrotor. Only some initial states converge to the path. Which ones?
Where the controller is defined, the transverse dynamics are exactly linear. The state-constrained Zubov value function then splits into a known quadratic cost-to-go plus a safety value function, and only the latter is learned.
An estimate with an inner layer verified by interval branch and bound and an outer layer tested on samples. It covers on average – and – of the simulated region of attraction on the unicycle and – on the planar quadrotor.
The question
Within each animation, every vehicle runs the same controller. While a vehicle stays admissible, its transverse error obeys linear, exponentially stable dynamics. Some vehicles reach the singular set or a state constraint first, and then the controller fails. The constrained region of attraction is the set of starting states that never reach the singular set or leave the state constraints.
Fifty vehicles start heading north at speed 1. Those in the simulated region swirl onto the circle. Red marks show runs that leave the state constraints or reach the centre of the circle, where the path coordinate is undefined.
The same vehicle on a periodic path. On this slice the only failures are exits through the state constraint on x₂.
Zero thrust is singular. Blue vehicles start in the estimate and orbit the loop. The others drive their thrust to zero and fail.
Launch a vehicle
Click anywhere on the map to launch a unicycle with the heading and speed below. The page integrates the true closed loop in your browser. Watch the three traces: while the vehicle stays admissible, the transverse error obeys linear, exponentially stable dynamics, so the only way to fail is to leave the admissible set.
Map: the slice of the state space at the chosen heading and speed. Aqua is the simulated region of attraction, blue our estimate (union of four seeds), orange the verified inner layer.
The idea
The state-constrained Zubov value function of Grüne and Zidani is finite exactly on the constrained region of attraction. On the admissible set the error coordinates obey ż = Az, so the running cost integrates to a quadratic, and the value function splits in two:
The first term is the quadratic cost-to-go. It is exact, and it decreases strictly along every trajectory in the admissible set until the trajectory reaches the path. The second is a safety value function: the worst future margin to the constraints. It is the only part a network has to learn, and it is learned by fitted value iteration from short trajectory segments, without infinite-horizon trajectory labels. In the traces above, the orange part always falls. The blue part is positive only while the vehicle will later approach a constraint, and it never grows along a run.
A run that passes close to a state constraint, the face x₂ = −2. The bar shows V at the current state, split into the known part and the part a network learns.
Two layers
The estimate is a sublevel set {v < c₂} of the learned function, joined to a sublevel set of the quadratic cost-to-go that interval branch and bound verifies. The level c₂ is set by counterexamples, states where v fails to decrease. Drag the level well past the selected value and watch red states appear: the estimate then contains states outside the simulated region of attraction.
One seed, one slice of the four-dimensional state space at a time. Coverage here is measured on the 121 × 121 slice grid, so it differs from the numbers in the paper, which are measured on the whole box.
What a million samples buy
The direct test of unsafe states draws N independent samples of the box and simulates each sample in the estimate until it reaches the verified inner layer. If k of them are unsafe, then with confidence 1 − 10⁻⁶ and up to the integration error, the unsafe states occupy at most ε of the box, where ε is the one-sided Clopper–Pearson bound.
In the paper, with 10⁶ box samples per case, the test finds k = –, – and – in the unions of four estimates of the circle, the sinusoid and the quadrotor, giving ε ≤ –, – and –.
largest share of the box that can still fail
The marker shows ε on a log scale of the share of the box.
What holds, and how
On the admissible set, the constrained region of attraction of a TFL controller is its largest forward invariant subset, and the value function splits into the quadratic cost-to-go and the safety value function.
A sublevel set of the quadratic cost-to-go, checked by interval branch and bound in floating-point arithmetic with outward rounding. It does not depend on the network, and {v < c₁} lies inside it for every candidate.
The learned sublevel set, tested with 10⁶ samples. A scenario bound limits the missed counterexamples, and a direct test of unsafe states limits the states it wrongly includes.
Results
Coverage of the simulated region of attraction, mean over four seeds. Every method passes the same level selection and final test. The no-decomposition baseline uses the same network, data, losses and fine-tuning without the quadratic cost-to-go in the value function. Toggle methods, and hover a marker for its value.
The hollow marker is the exact value function cut at the level of the candidate, which shows how much the level alone allows. The quadrotor is the hardest case: counterexamples hold its level low, and the bound V ≥ γqex alone then excludes – of its simulated region of attraction.
One algorithm
The algorithm trains seed 0 with a long-step and a short-step setting, and keeps, for all seeds, the setting whose estimate holds the larger share of box samples before fine-tuning. On the TFL closed loops, the long step is kept on –.
Beyond TFL, the same algorithm applies to closed loops with an exponentially stable equilibrium, with the quadratic form of the linearization in place of the exact cost-to-go. The decomposition is then no longer exact, yet on one seed of five benchmarks the estimates cover between – and – of the simulated region of attraction.
Seed 0. Share: share of box samples in the estimate before fine-tuning, which decides the setting. Coverage: share of the simulated region of attraction covered with the kept setting after fine-tuning.
Reproduce
# clone and set up git clone https://github.com/gradslab/roa_tfl.git cd roa_tfl bash reproduce.sh setup # every table and number of the paper, compared with the paper bash reproduce.sh tables # Figure 1 and the paper PDF bash reproduce.sh figure bash reproduce.sh paper # the full computation from scratch (three to four days on a laptop CPU) bash reproduce.sh all
| tfl/ | the three TFL cases: training, interval verification, level selection, final test, fine-tuning, baselines, and the generator of every number |
| benchmarks/ | five closed loops with an exponentially stable equilibrium, trained on seed 0 with both step settings |
| paper/ | the paper sources and the reference copy of every generated table |
| render/, web/ | the animations and the data of this page |
Cite
@unpublished{malik2027constrained,
author = {Abdullah Waqar Malik and Hamza Tariq and Adeel Akhtar},
title = {Learning Regions of Attraction for Transverse Feedback
Linearization under Controller Singularities},
note = {Under review},
year = {2027}
}