State-Dependent Conformal Perception Bounds
for Neuro-Symbolic Verification
Thomas Waite1
, Yuang Geng2
, Trevor Turnquist2
, Ivan Ruchkin2
*, Radoslav Ivanov1
*
1
Rensselaer Polytechnic Institute
2
University of Florida
* Co-last authors: equal contribution in supervision
July 21, 2025 — Zagreb, Croatia — co-located with CAV’25
3rd TACPS workshop
Supported by the NSF CPS Grants CCF-2403615 and CCF-2403616
• TEA Lab’s mission: make autonomous CPS safer and more trustworthy through
rigorous engineering methods with strong guarantees
• Across many domains: autonomous driving, mobile robots, UAVs/UUVs
• Research areas: formal methods, AI/ML, CPS, robotics
• Theory → techniques/algorithms/tools → simulations+physical deployment
TEA Lab: Trustworthy Engineered Autonomy
tea.ece.ufl.edu
2
Research platform:
small-scale racing cars
ivan.ece.ufl.edu
iruchkin@ece.ufl.edu
• 2006–2011: Undergrad @ Moscow State University
• Moscow, Russia — applied math & computer science
• Industry and government roles
• AFRL VFRP, NASA JPL, CMU SEI, Google Summer of Code, …
• 2011–2018: PhD @ Carnegie Mellon University
• Pittsburgh, PA — software engineering & formal methods
Ivan Ruchkin: Brief Bio
3
• 2018–2022: Postdoc @ University of Pennsylvania
• Philadelphia, PA — guarantees for learning-enabled systems
• 2022–present: Assistant professor @ University of Florida
• Gainesville, FL — trustworthy methods for autonomy
Motivation: Safety for Neural CPS
• Modern cyber-physical systems use advanced neural perception & control,
but lack safety guarantees
• Existing approaches attempt to bridge neural perception & verification, but
suffer from excessive conservatism and confidence decay over time
4
Perception-driven Systems
Major Question:
How do we provide tight,
high-confidence safety guarantees for
perception-driven systems?
5
Original Blurred
[ T. Waite, Y. Geng, T. Turnquist, I. Ruchkin, R. Ivanov, “State-Dependent Conformal
Perception Bounds for Neuro-Symbolic Verification of Autonomous Systems”, NeuS 2025 ]
High-Confidence Reachability Problem
6
Related Work
• Vanilla conformal prediction (CP) leads to overly conservative guarantees
without considering closed-loop dynamics
[ V. Vovk, “Algorithmic Learning in a Random World”, 2005 ]
• Time-based CP exploits the dependency of error on time — but not state
[ M. Cleaveland, I. Lee, G. Pappas, L. Lindemann, “Conformal prediction regions for time series using
linear complementarity programming”, AAAI 2024 ]
• Other works combine CP with reachability verification — but in a different
and/or handcrafted pattern
[ Y. Geng, J. Baldauf, S. Dutta, C. Huang, I. Ruchkin, “Bridging Dimensions: Confident Reachability for
High-Dimensional Controllers”, FM 2024 ]
• Our insights: (a) perception error can also be heteroskedastic in state, and
(b) the knowledge of system dynamics can reduce conservatism
7
Background: Scalar Conformal Prediction
8
Given:
• Calibration dataset
• Desired confidence
Provides:
➔ Upper bound for the next sample:
Details: [ Lars Lindemann’s & Jyo Deshmukh’s tutorial
“Formal Verification and Control with Conformal Prediction” ]
State-Dependent Conformal Prediction Problem
9
Approach Overview
10
NEURAL NEURO-SYMBOLIC
SYMBOLIC
Step 2: High-Confidence Reachability
11
Approach Overview
12
Step 1: Searching for State Regions
13
Solution recipe:
1. Define a loss function over perception error bounds
2. Synthesize boundaries of state regions
3. Obtain an upper error bound for each with confidence
Solve with gradient-free optimization:
Simulated Annealing (SA) or
Genetic Algorithm (GA)
Experience Loss (EL):
Experience Time-Decay Loss (EL):
14
(a) Original (b) Low-Contrast (c) High-Contrast (d) Blur
Case Study: Mountain Car
Our Error Bounds Are Tighter
15
• Perception error is heteroskedastic over both space and time
• Trade-off between low- and high-error regions
Our Reachsets Are Tighter
16
• Our method (GA+ETDL) for 7 regions vs. the time-based method
Our Method is Slower But Less Conservative
17
Limitations
1. Inflexible confidence: same for all regions
2. Partitioning along only one state dimension
3. Unclear what partitioning is truly the best for reachability verification
18
Summary
19
[ T. Waite, Y. Geng, T. Turnquist, I. Ruchkin, R. Ivanov, “State-Dependent Conformal
Perception Bounds for Neuro-Symbolic Verification of Autonomous Systems”, NeuS 2025 ]
1. Optimized conformal bounds
over state regions
2. High-confidence verification of perception-driven systems
3. Evaluation on a mountain car with visual distribution shift
NEURAL NEURO-SYMBOLIC
SYMBOLIC

State-Dependent Conformal Perception Bounds for Neuro-Symbolic Verification

  • 1.
    State-Dependent Conformal PerceptionBounds for Neuro-Symbolic Verification Thomas Waite1 , Yuang Geng2 , Trevor Turnquist2 , Ivan Ruchkin2 *, Radoslav Ivanov1 * 1 Rensselaer Polytechnic Institute 2 University of Florida * Co-last authors: equal contribution in supervision July 21, 2025 — Zagreb, Croatia — co-located with CAV’25 3rd TACPS workshop Supported by the NSF CPS Grants CCF-2403615 and CCF-2403616
  • 2.
    • TEA Lab’smission: make autonomous CPS safer and more trustworthy through rigorous engineering methods with strong guarantees • Across many domains: autonomous driving, mobile robots, UAVs/UUVs • Research areas: formal methods, AI/ML, CPS, robotics • Theory → techniques/algorithms/tools → simulations+physical deployment TEA Lab: Trustworthy Engineered Autonomy tea.ece.ufl.edu 2 Research platform: small-scale racing cars ivan.ece.ufl.edu iruchkin@ece.ufl.edu
  • 3.
    • 2006–2011: Undergrad@ Moscow State University • Moscow, Russia — applied math & computer science • Industry and government roles • AFRL VFRP, NASA JPL, CMU SEI, Google Summer of Code, … • 2011–2018: PhD @ Carnegie Mellon University • Pittsburgh, PA — software engineering & formal methods Ivan Ruchkin: Brief Bio 3 • 2018–2022: Postdoc @ University of Pennsylvania • Philadelphia, PA — guarantees for learning-enabled systems • 2022–present: Assistant professor @ University of Florida • Gainesville, FL — trustworthy methods for autonomy
  • 4.
    Motivation: Safety forNeural CPS • Modern cyber-physical systems use advanced neural perception & control, but lack safety guarantees • Existing approaches attempt to bridge neural perception & verification, but suffer from excessive conservatism and confidence decay over time 4
  • 5.
    Perception-driven Systems Major Question: Howdo we provide tight, high-confidence safety guarantees for perception-driven systems? 5 Original Blurred [ T. Waite, Y. Geng, T. Turnquist, I. Ruchkin, R. Ivanov, “State-Dependent Conformal Perception Bounds for Neuro-Symbolic Verification of Autonomous Systems”, NeuS 2025 ]
  • 6.
  • 7.
    Related Work • Vanillaconformal prediction (CP) leads to overly conservative guarantees without considering closed-loop dynamics [ V. Vovk, “Algorithmic Learning in a Random World”, 2005 ] • Time-based CP exploits the dependency of error on time — but not state [ M. Cleaveland, I. Lee, G. Pappas, L. Lindemann, “Conformal prediction regions for time series using linear complementarity programming”, AAAI 2024 ] • Other works combine CP with reachability verification — but in a different and/or handcrafted pattern [ Y. Geng, J. Baldauf, S. Dutta, C. Huang, I. Ruchkin, “Bridging Dimensions: Confident Reachability for High-Dimensional Controllers”, FM 2024 ] • Our insights: (a) perception error can also be heteroskedastic in state, and (b) the knowledge of system dynamics can reduce conservatism 7
  • 8.
    Background: Scalar ConformalPrediction 8 Given: • Calibration dataset • Desired confidence Provides: ➔ Upper bound for the next sample: Details: [ Lars Lindemann’s & Jyo Deshmukh’s tutorial “Formal Verification and Control with Conformal Prediction” ]
  • 9.
  • 10.
  • 11.
    Step 2: High-ConfidenceReachability 11
  • 12.
  • 13.
    Step 1: Searchingfor State Regions 13 Solution recipe: 1. Define a loss function over perception error bounds 2. Synthesize boundaries of state regions 3. Obtain an upper error bound for each with confidence Solve with gradient-free optimization: Simulated Annealing (SA) or Genetic Algorithm (GA) Experience Loss (EL): Experience Time-Decay Loss (EL):
  • 14.
    14 (a) Original (b)Low-Contrast (c) High-Contrast (d) Blur Case Study: Mountain Car
  • 15.
    Our Error BoundsAre Tighter 15 • Perception error is heteroskedastic over both space and time • Trade-off between low- and high-error regions
  • 16.
    Our Reachsets AreTighter 16 • Our method (GA+ETDL) for 7 regions vs. the time-based method
  • 17.
    Our Method isSlower But Less Conservative 17
  • 18.
    Limitations 1. Inflexible confidence:same for all regions 2. Partitioning along only one state dimension 3. Unclear what partitioning is truly the best for reachability verification 18
  • 19.
    Summary 19 [ T. Waite,Y. Geng, T. Turnquist, I. Ruchkin, R. Ivanov, “State-Dependent Conformal Perception Bounds for Neuro-Symbolic Verification of Autonomous Systems”, NeuS 2025 ] 1. Optimized conformal bounds over state regions 2. High-confidence verification of perception-driven systems 3. Evaluation on a mountain car with visual distribution shift NEURAL NEURO-SYMBOLIC SYMBOLIC