Conformal Controllers: Rigorous Safety & Robustness

• Conformal controllers are control systems that use conformal prediction to deliver rigorous statistical guarantees on safety, performance, and stability.
• They construct nonparametric uncertainty sets from calibration data to tighten constraints and control risk under model uncertainty.
• Integrated into applications like robotics and navigation, they balance aggressive control policies with adaptive safety filters and online risk calibration.

A conformal controller is a control-theoretic design or runtime synthesis method that leverages conformal prediction or conformal mapping to obtain rigorous statistical or geometric guarantees on safety, performance, or stability. The conformal prediction framework is data-driven and distribution-free, providing error bounds, probabilistic invariant sets, or calibrated uncertainty margins with finite-sample, nonparametric coverage guarantees. Approaches span robust and optimal control of linear and nonlinear dynamical systems, safety filters for learning-based robots, risk-constrained policy deployment, black-box safety barrier adaptation, and geometric robot navigation. The term “conformal controller” is used for any control system or policy augmented by conformal inference at design or run time.

1. Conformal Prediction Principles in Control

Conformal prediction provides valid prediction sets or bands around estimates made by learned predictors, calibrated so that the true outcome lies within the set with a user-specified coverage level $1-\alpha$. In control, conformal prediction is applied in several contexts:

• Model uncertainty quantification: Envelope learned model predictions with conformal bounds, yielding finite-sample, distribution-free coverage on the realized error, e.g., model residual norms $r_i = \|y_i-f(x_i)\|$ on held-out data; the $(1-\alpha)$ quantile determines the set (Nath et al., 3 Jan 2026, Wei et al., 18 Jul 2025).
• Constraint tightening and safety regions: Inflate reachable or safe sets by conformal prediction radii to account for modeling uncertainty or disturbances, guaranteeing that safety constraints are met with probability at least $1-\alpha$ (Ma et al., 3 Feb 2026, Vlahakis et al., 2024, Tabbara et al., 11 Nov 2025).
• Risk control in policies and decision rules: Adjust policy conservatism or constraint slack online by tracking empirical violations and calibrating to hold average risk at most $\alpha$ via a single scalar update (Lekeufack et al., 2023, Huriot et al., 2024, Prinster et al., 2 Mar 2026).

The core theoretical guarantee is marginal coverage under mild exchangeability conditions or even adversarial conditions in online settings.

2. Data-Driven Robust Control with Conformal Uncertainty Sets

Distribution-free robust control is obtained by constructing conformal-prediction-based uncertainty sets around learned system or disturbance models:

• Linear systems (LQR/MPC): Given a learned dynamics predictor $\widehat{C} = f(\theta)$, conformal calibration on held-out samples yields a set $$\mathcal{U}(\theta) = \{C: \|C - f(\theta)\|_2 \le \widehat{q}\}$$, where $\widehat{q}$ is the $(1-\alpha)$ quantile. The resulting robust control problem is

$$K_{\rm rob}(\theta) = \arg\min_K \max_{C \in \mathcal{U}(\theta)} J(K, C)$$

with high-probability coverage on regret bounds (Patel et al., 2024).

• Nonlinear continuous-time systems: A nominal contraction-based design is rendered robust by inflating the contraction test with the conformal-quantile radius $r_i = \|y_i-f(x_i)\|$0 of the predictor error $r_i = \|y_i-f(x_i)\|$1, guaranteeing exponential boundedness of tracking error with probability $r_i = \|y_i-f(x_i)\|$2 (Wei et al., 18 Jul 2025).
• Koopman-based embeddings: Conformal regions for the lifted residual $r_i = \|y_i-f(x_i)\|$3 in $r_i = \|y_i-f(x_i)\|$4 are constructed, and tracking error bounds are propagated via contraction metrics from lifted to original coordinates (Hirano et al., 23 Mar 2026, Nath et al., 3 Jan 2026).
• Joint chance constraints: Finite-horizon or trajectory-wide conformal bands are computed on closed-loop error trajectories $r_i = \|y_i-f(x_i)\|$5, defining a prediction region (PR) as a Euclidean ball or ellipsoid. This allows conversion of stochastic constraints to deterministic tightenings with guaranteed feasibility (Vlahakis et al., 2024).

These approaches yield probabilistically safe controllers with coverage guarantees uniform in the sample size and without needing parametric distributional assumptions.

3. Safe Filters and Switching via Conformal Certification

Conformal safety filters mediate between aggressive (performance-seeking) and conservative (risk-averse) controllers, certifying at each step whether a nominal action is safe relative to conformal bounds:

• HJ Reachability and RL: Learned value functions and policies are subject to split-conformal correction, deriving lower bounds on the true safety return. At runtime, a switched controller uses the nominal policy if the conformal bound certifies safety, otherwise reverts to a guaranteeable safe mode (Tabbara et al., 11 Nov 2025).
• RL Policy Wrapping: Pre-trained RL policies are wrapped by a conformal predictive safety filter (CPSF) which solves a receding-horizon MPC, ensuring planned agent paths avoid conformal–quantile inflated prediction balls of other agents, guaranteeing a user-specified collision avoidance probability (Strawn et al., 2023).
• Ensemble methods: Ensembles of independently trained safety filters, each with its own conformal calibration, further tighten safety guarantees by selecting the best (most conservative) member at each step (Tabbara et al., 11 Nov 2025).

Empirical evaluation demonstrates significant reduction in safety violation rates compared to standard or fixed-margin baselines across robotic, manufacturing, and trading tasks.

4. Online and Adaptive Approaches: Decision-Theoretic and Policy Conformal Controllers

A distinct class consists of online controllers that directly calibrate the tradeoff between safety and performance via a scalar parameter, adapting in response to observed losses:

• Conformal Decision Theory (CDT): Maintains a risk-control variable $r_i = \|y_i-f(x_i)\|$6 in one-dimensional policy or constraint parameterizations, updated online via $r_i = \|y_i-f(x_i)\|$7, ensuring empirical risk $r_i = \|y_i-f(x_i)\|$8 under only an "eventual safety" assumption (Lekeufack et al., 2023).
• Conformal Policy Control (CPC): Given a reference policy $r_i = \|y_i-f(x_i)\|$9 (with risk guarantee) and a candidate $(1-\alpha)$0, CPC uses conformal calibration to choose a likelihood-ratio clip threshold $(1-\alpha)$1, forming the policy $(1-\alpha)$2, which provably controls average constraint violation with finite-sample guarantees, including for non-monotonic and high-dimensional losses (Prinster et al., 2 Mar 2026).
• Decentralized adaptation: In multi-agent or black-box prediction settings, a single conformal variable (e.g., a constraint slack $(1-\alpha)$3) is updated online to match an empirical violation rate, guaranteeing an average safety gap that holds uniformly over time (Huriot et al., 2024).

These frameworks abstract away the details of prediction sets and focus on empirical risk, scalability, and modularity for arbitrary black-box controllers or policies.

5. Geometric and Conformal-Mapping Based Controller Design

Beyond statistical conformal prediction, conformal geometric mappings are utilized to facilitate feedback design, particularly in navigation and motion-planning:

• Conformal navigation transformations: Complex robot workspaces with arbitrary obstacle shape are conformally mapped onto a standard multiply-connected "sphere world." Gradient navigation functions are designed in this canonical domain, then pulled back via the conformal map to yield controllers with guaranteed obstacle avoidance and convergence in the original geometry (Fan et al., 2022).
• Fractional-order PID tuning via conformal mapping: FOPID controllers are interpreted via a conformal map $(1-\alpha)$4, where non-integer order zeros become regular zeros in $(1-\alpha)$5-space. Varying $(1-\alpha)$6 shifts the zeros/poles along an “M-curve” in the complex plane, providing a systematic means to tune PID gains for increased damping and reduced control effort while preserving dominant pole locations (Saha et al., 2012).

Geometric conformal mapping thus serves as an analytic tool for both transferring control design between spaces and synthesizing controllers that combine robustness and performance.

6. Empirical Results, Benchmarks, and Domains of Application

Across the literature, conformal controllers have been validated in a range of challenging domains:

Domain	Typical Conformal Use	Key Guarantees
RL & Robotics	Predictive safety filters, robust reachability	Finite-sample, distribution-free collision avoidance (Strawn et al., 2023, Ma et al., 3 Feb 2026)
Linear Control	Robust LQR/MPC under uncertain models	Probabilistic regret/control cost bounds (Patel et al., 2024, Vlahakis et al., 2024)
Nonlinear Control	Contraction-based tracking, Koopman & neural lifts	Exponential error/tube bounds (Wei et al., 18 Jul 2025, Hirano et al., 23 Mar 2026, Nath et al., 3 Jan 2026)
Policy Deployment	Online risk calibration, policy interpolation	User-declared average constraint violation, model agnostic (Prinster et al., 2 Mar 2026, Lekeufack et al., 2023)
Navigation	Geometric conformal transformations	Almost-global convergence and obstacle avoidance (Fan et al., 2022)

Experimental results consistently show that conformal controller designs either match or outperform margin-based or adversarial robustification, providing tighter confidence regions, lower conservatism, and uniformly controlled violation rates, even under context shift, dynamics model errors, or adversarial inputs.

7. Theoretical and Practical Implications

The key theoretical contribution of the conformal controller paradigm is the unification of machine learning uncertainty quantification and control-theoretic safety verification under a rigorous, data-driven, and nonparametric framework. The main features are:

• Finite-sample, distribution-free calibration: Guarantees hold regardless of the distribution of errors, model residuals, or disturbances, provided held-out data are suitably exchangeable or i.i.d.
• No reliance on parametric noise models: In contrast to Gaussian or bounded-norm disturbance assumptions, conformal methods adapt to empirical error structure.
• Modular, easily integrated algorithms: Most conformal controller implementations wrap existing model-based or model-free control pipelines, requiring only additional calibration data and quantile-based adaptation.
• Provable, quantitative risk–performance trade-offs: User-specified risk levels map directly to error bars, constraint tightenings, or policy clip thresholds, with no hyperparameter tuning beyond the chosen $(1-\alpha)$7.
• General applicability: Conformal controllers have been realized for linear, nonlinear, stochastic, and hybrid systems, in both batch and online settings, and for both feedback and policy optimization paradigms.

A plausible implication is that data-driven autonomy in safety-critical environments can increasingly abandon hand-tuned margins or overly conservative worst-case reasoning in favor of calibrated, empirically verifiable guarantees supplied by conformal controllers. Future directions include high-dimensional, context-adaptive calibration, integration with model-based reinforcement learning, and scalable certification in multi-agent or nonstationary settings.