Researchers learn safer activation zones for path-following controllers
An arXiv paper proposes a hybrid method that estimates where a transverse feedback linearization controller can be switched on safely, combining a verified inner estimate with a learned outer one.
In plain words
Some robots follow paths using a controller that stops working at certain states, like a unicycle at standstill. The authors compute which starting positions let the robot reach the path safely. Their trick: calculate part of the answer with equations and learn the rest with a value-iteration algorithm.
What was announced
Transverse feedback linearization steers a system onto a desired path by exactly linearizing its error dynamics, but the controller breaks down on a singular set, for instance when a unicycle stands still or a quadrotor produces no thrust. The paper studies the constrained region of attraction: the starting states from which the closed loop reaches the path while respecting state and regularity limits. Knowing this set tells operators where activation is safe. Using the known linear error dynamics, the authors prove the region equals the largest forward invariant subset of the admissible states, and they split its Zubov value function into an analytic quadratic part plus a safety term that a learning scheme must supply.
Only the safety term is learned, through fitted value iteration that avoids infinite-horizon trajectory labels. The final estimate has two layers: an inner one checked with interval branch-and-bound, and an outer one evaluated on samples. In tests on a unicycle tracking a circle and a sinusoid, plus a planar quadrotor, the method covered on average 72.0%, 83.4% and 19.1% of the simulated region of attraction. A plain quadratic baseline and two physics-informed neural solvers reached at most 11.3%, 4.0% and 0.0% on average. The authors note code and animations are published on a project page. These figures come from the paper's own simulations and have not been independently confirmed.
What it does not show
- Coverage numbers rest on the authors' own simulations of a unicycle and a planar quadrotor; no independent replication is reported.
- The source is an abstract, so full method details, assumptions and proofs are not available for verification.
- The outer layer is only sample-tested, so its guarantees are weaker than the interval-verified inner layer.
For security and IT teams
- Scope
- Results apply to transverse feedback linearization controllers whose singular set, such as zero speed or zero thrust, limits where they may be activated.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.