Workshop Scope
Robots are increasingly powered by learned perception, prediction, and decision models. Recent progress in reinforcement learning and robotic foundation models has further amplified this trend. However, guaranteeing safety and verifying the behaviors of learned models is difficult, due to, for example, the use of deep neural networks and the potential for hallucinations. Reliable deployment of robots requires explicit guarantees about safety and task satisfaction. Formal methods—a set of rigorous techniques for specifying, analyzing, and verifying systems—offer a framework for considering such problems.
This workshop focuses on two complementary viewpoints at the interface of learning and formal methods: learning for formal methods and learning with formal methods. Learning for formal methods studies how data-driven techniques can provide the missing ingredients that formal synthesis and verification often assume, including specifications and constraints inferred from demonstrations, preferences, or language, as well as abstractions, uncertainty sets, and conservative models that enable provable reasoning. Learning with formal methods studies how formal tools can guide and constrain learning, for example, through specification-aware training, verification, and certification of learned components, and runtime monitoring and shielding.
We bring together researchers from various disciplines—including robot learning, formal methods, planning, controls, safety, and multi-robot systems—to share past work, recent advances, and open problems in this exciting research area. The workshop will consist of a diverse set of invited talks and poster sessions. We hope to identify new research directions, open problems, and interdisciplinary perspectives that bridge these two fields and lead to new advances in robotics.
The workshop is organized along two complementary directions:
- Learning for formal methods: specification/constraint inference, learning abstractions and models, uncertainty descriptions for synthesis.
- Learning with formal methods: formally grounded learning objectives, verification and certification of learned components, runtime assurance (monitors, shields, fallback controllers).
Topics of interest include, but are not limited to:
- Understanding specifications and constraints: learning from demonstrations, preferences, or language; learning abstractions (predicates, automata, contracts); learning models and uncertainty sets for synthesis.
- World-model training and verification: specification-aware training; verification and certification of learned policies and components; formal runtime assurance.
- World models in robotics: safe reinforcement learning, certifiable perception and prediction, verified learned components, multi-robot systems.
- Benchmarks, datasets, and demonstrations for learning and formal methods in robotics.
Invited Speakers & Panelist
(acceptance order)
Organizing Team
* Corresponding organizers
Schedule & Speakers
October 1, 2026 · Pittsburgh, PA
| Time | Activity |
|---|---|
| 13:30 – 13:35 | Introduction |
| 13:35 – 13:55 | Invited talk 1: Ahmed Qureshi — Formal Specifications as Inductive Bias for Robot Policy Learning and Control |
| 13:55 – 14:15 | Invited talk 2: Chuchu Fan — Scalable Learning for Safe and Task-Correct Autonomy |
| 14:15 – 14:35 | Junior research talk 1: Yuwei Wu — Abstractions for Learning and Formal Reasoning in Trustworthy Robot Autonomy |
| 14:35 – 15:05 | Invited talk 3: Sheila McIlraith — Exploiting Formal Languages in Robot Policy Learning: Faster, Safer, and Better |
| 15:05 – 15:25 | Junior research talk 2: Disha Kamale — Closing the Abstraction Gap: Adaptive Abstractions for Provably Safe Robot Planning |
| 15:25 – 16:20 | Poster session and coffee break |
| 16:20 – 16:50 | Invited talk 4: Yiannis Kantaros — Assured Language-driven Autonomy: Uncertainty-aware Translation of Natural Language to Formal Specifications |
| 16:50 – 17:20 | Invited talk 5: Glenn Chou — Planning Safely with Learned Dynamics Beyond the Training Data |
| 17:20 – 17:30 | Closing remarks and awards |
Accepted Papers
The following papers were accepted to the workshop.
- Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions
- The Hidden Alphabet of Robot Tasks: A Partition-Lattice Theory of Identifiability and Safe Interventions
- A Robot Can Learn Only What It Can Discriminate: Automata–Chernoff Duality and Instance-Optimal Identification of Temporal Objectives
- Withdrawal Constrained Transfer for Revising Rollout Support in World Model Planning
- From Emergent Messages to Formal Interfaces: A Finite-Sample Study of Communication-Conditioned Behavioral Contracts
- GRAND: Guidance, Rebalancing, and Assignment for Networked Dispatch in Multi-Agent Path Finding
- Learning Object-Centric Barrier Functions for Contact-Rich Interaction in Clutter
- From LLM-Generated Specifications to Learned Quadruped Locomotion
- Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
- Abstractions for Learning and Formal Reasoning in Trustworthy Robot Autonomy
- ASPIRE: Decoupling Perception and Reasoning in Vision-Language Models with Answer Set Programming
- Logic-VLA: A Temporal Logic Conditioned Vision-Language-Action Model
- PDDL-ART: Autonomous Symbolic Abstraction From Demonstration For Long-Horizon Robotic Manipulation Using Vision-Language Models
- Runtime-Constrained Residual Execution for Sequential Dual-Sided Grasping
- IFG: Internet-Scale Guidance for Functional Grasping Generation
- Closing the Abstraction Gap: Adaptive Abstractions for Provably Safe Robot Planning Oral
- Design and Learning of Spatiotemporal Metrics for Trajectory Optimization in Dynamic Environments
- Inference-Time Robot Behavior Steering through Physically-Aware Reconfiguration of Task-Structure
- Discovering Formal Options with Contrastive Policy Composition
Call for Papers
We invite contributions centered around the following two perspectives:
- Learning for formal methods: specification/constraint inference (from demos, feedback, language), learning abstractions and compositional interfaces (predicates, automata, contracts), and learning models and uncertainty descriptions for synthesis (calibration, conformal bounds, conservative dynamics/environment models).
- Learning with formal methods: formally grounded learning objectives (logic- or robustness-based losses, constrained learning/RL), verification and certification of learned policies and components (NN verification, certified robustness, barrier/reachability certificates), and formal runtime assurance (monitors, shields, fallback controllers, falsification and counterexample-guided data collection).
A regular paper submission should be 4–8 pages in length (including references). Late-breaking results submissions should be 2 pages (including references). Submissions should concisely describe the core idea and results, highlight novelty, and clearly position the work within one or both workshop perspectives (learning for formal methods, learning with formal methods). All papers must be submitted in PDF format and follow the standard IEEE conference formatting guidelines (templates are available on the IEEE website).
The workshop is non-archival and dual submissions are allowed. Submissions may present preliminary and/or completed work. Reviewing is single-blind.
Submissions will be reviewed by a Program Committee composed of the organizers, invited speakers, and additional experts in the area. Reviews will consider the following criteria.
- Relevance and positioning: Does the submission clearly connect to learning for formal methods (e.g., learning specifications, abstractions, models, uncertainty descriptions) and/or learning with formal methods (e.g., formally grounded learning objectives, certification, runtime assurance)?
- Technical clarity: Are the problem setting, assumptions, and evaluation protocol stated clearly enough to assess the claims?
- Methodological soundness: Are the proposed methods well motivated and technically credible, including how uncertainty, ambiguity, or robustness is handled when applicable?
- Novelty and potential impact: What is new relative to prior work, and how likely is the approach to influence future research or enable stronger guarantees in learning-enabled robotics?
- Insights and outlook: Does the paper surface limitations, open problems, or concrete research questions that can shape follow-on work by the community?
Accepted papers will be posted on the workshop website, and at least one author of each accepted paper will be invited to present in the workshop poster session.
Submission: Submit via OpenReview, selecting the Paper Submission track.
Call for Talk Proposals
We invite junior researchers, who are either close to completing their PhD or are recent graduates, to share their PhD research work and research vision on learning + formal methods at our workshop as a 15-min talk.
Applicants must have either defended their PhD thesis after May 2024 or must be in their 3+ years of PhD study. Applicants are invited to submit a talk proposal in the form of an extended abstract of up to 2 pages (excluding references) summarizing their PhD research on a topic of interest to the workshop. The workshop is non-archival and dual submissions are allowed. Reviewing is single-blind.
One proposal will be selected for presentation; others may be considered for the poster session.
Submission: Submit via OpenReview, selecting the Talk Proposal Submission track.