Authors
Hadi Jahanshahi, Zheng Hong Zhu
Published in
PloS one. Volume 21. Issue 11. Pages e0352196. Epub Nov 10, 2026.
Abstract
Characterizing the performance limits of physical systems from experimental data is essential for engineering decision-making, design optimization, and reliability analysis, yet conventional regression models estimate average behavior rather than the boundary of what is achievable. This paper develops a data-driven methodology for discovering nonlinear performance boundaries from experimental observations. The problem is formulated as extreme conditional quantile estimation. A nonparametric binned percentile estimator recovers low-quantile boundary points, and the Bayesian Information Criterion selects an interpretable closed-form law from a library of physically motivated families. Non-asymptotic guarantees are stated in the fixed-design regime the experiments occupy. The method is validated on two real robotic datasets. For UR10e trajectory tracking (30,600 observations), it discovers a saturation boundary E=7.74(1-e-13.27S) mm (R2=0.962; bootstrap 95% CI for Emax of [7.05, 9.10] mm) and separates it from a steady-state envelope obeying the lag law e=tlag·v with tlag=65 ms (R2=0.997) once start-up transients are excluded. For planar pushing (40,378 pushes from the MIT MCube dataset), it discovers a power-law boundary with exponent α=2.02 (95% CI [1.88, 2.17]), indicating that the root-mean-square force fluctuation grows linearly with speed, consistent with momentum-transfer scaling at the breakdown of the quasi-static contact assumption. Family selection is stable across quantile levels and bootstrap resampling schemes, recovers the correct family in every synthetic replicate at the experimental configurations for the two family classes discovered here, and is confirmed by a prospectively specified experiment. Six of six extrapolation predictions at never-tested speeds landed within 1.2 mm, and a repeat of the protocol reproduced the per-episode 2nd percentiles to a mean absolute deviation of 0.039 mm. Compared with quantile regression, Gaussian process, support vector, and frontier-estimation baselines, the framework provides the tightest interpretable boundary estimates while approaching the target validity level, and it extrapolates where nonparametric envelopes cannot.
PMID:
43007980
Bibliographic data and abstract were imported from PubMed on 11 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 19
- Comments 0