Workshop on Reliability In Planning and Learning (RIPL; joint with HSRL)

Aim and Scope of the Workshop

Learning is the dominating trend in AI at this time, achieving (among others) unprecedented versatility and scalability in many forms of sequential decision making. Given the opaque nature of ML models and the lack of inherent guarantees, reliability is a key concern, prominently including safety, robustness, and fairness in various forms, but possibly other concerns as well. Arguably, this is indeed one of the grand challenges in AI for the foreseeable future. Research on this challenge is widespread across the AI community and beyond. Research topics relevant to ICAPS include, for example, safe and high-stakes reinforcement learning, quality assurance for LLM-generated plans or planning models, as well as stress testing and formal verification of learned action policies. The mission of this workshop is to represent this important topic space at ICAPS, providing a joint discussion forum, and gradually forming a sub-community, addressing any topic related to reliability issues in the use of ML methods for planning and scheduling purposes.

The first workshop of the RIPL series at ICAPS was held in 2022, and ran through 2024, then under the name RDDPS. The workshop was renamed to RIPL in 2025 reflecting a more inclusive scope. In the 2026 edition, RIPL is merged with another proposed workshop centering on high-stakes reinforcement learning, expanding our vision to high-stakes domains where traditional trial-and-error learning is infeasible and thus explicit world models and planning as strict guardrails for safe deployment are needed.

From a planning and scheduling perspective – and for sequential decision making in general – the importance of learning is manifested in two major kinds of technical artifacts that are rapidly gaining importance. First, planning models partially learned from data (such as a weather forecast in a model of flight actions), or generated by LLMs. Second, action-decision components learned from data, in particular action policies or planning-control knowledge for making decisions in dynamic environments (e.g., manufacturing processes under resource-availability and job-length fluctuations).

Reliability of data-driven artifacts, in particular ML classifier robustness and fairness, is one of the key research issues in other sub-areas of AI for quite some time already. Yet the topic has so far been scarcely addressed at ICAPS, whose focus in planning and learning has so far been mainly on plan-generation performance. The organizers of this workshop believe that this needs to change, as it is important that ICAPS contributes to address the reliable AI challenge. We furthermore believe that ICAPS is in a good position to make such a contribution, as the combination of symbolic and data-driven methods is a key avenue for obtaining reliable AI. The workshop aims at establishing an ICAPS sub-community focusing on this vision.

Schedule (Preliminary) – June 29, Room 6

  • Technical talks will be 15 minutes + 5 minutes for discussion

9:00 - 10:00 Session: Welcome and Invited Talk

  • Welcome and Open Remarks

  • Invited Talk

Siddharth Srivastava
Siddharth Srivastava ‐ The Return of Structure: Learning Symbolic World Models from Raw Data for Reliable Planning and RL

Abstract: Generative AI systems achieve impressive performance through scale, yet struggle with reliability, generalizability and data efficiency. This talk explores a complementary approach based on learning structure: from Conditional Abstraction Trees that autonomously discover useful abstractions for reinforcement learning to recent methods that learn symbolic world models directly from raw numeric observations. By combining the learning capabilities of modern machine learning with the compositional reasoning strengths of automated planning algorithms, these methods enable more reliable learning and data efficient behavior. I will discuss recent advances in learning symbolic abstractions from unstructured data, their role in planning and reinforcement learning, and why the return of structure may be essential for building reliable AI systems that operate beyond the training distribution.

10:00 - 10:30 Session 1: Multi-Agent Coordination under Uncertainty

  • Rolando Fernandez, Caleb Probine, Tyler Lee, Jeffrey Chen, Erez Karpas, Muhammad Arrasy Rahman, Peter Stone, and Ufuk Topcu – Social Laws for Multi-agent Coordination in Stochastic Environments (pdf, bibtex)

10:30 - 10:50 Coffee Break

10:50 - 12:30 Session 2: Neuro-Symbolic Policies and Their Verification

  • Marcel Schubert and Daniel Höller – Safety Verification of Additive Tree Ensemble Policies via Symbolic Search (pdf, bibtex)
  • Farid Musayev, Dominik Drexler, Daniel Gnad, and Jendrik Seipp – A Comparison of Sampling Strategies for Learning Policies (pdf, bibtex)
  • Victor Scherer Putrich, Tim Nico Bauerschmidt, Chaahat Jain, Songtuan Lin, Joerg Hoffmann, and Isabel Valera – Formula Neural Networks as Action Policies in Safety-Critical Planning: An Empirical Evaluation (SoCS pdf) (SoCS 2026 paper)
  • Dominik Drexler, Markus Fritzsche, Farid Musayev, and Simon Ståhlberg – Prompt, Prove, Patch: The Neuro-Symbolic Loop for General Policy Synthesis (pdf, bibtex)

12:30 - 14:00 Lunch Break

14:00 - 15:30 Session 3: Learning for Classical Planning

  • Mingyu Hao, Sylvie Thiebaux, and Felipe Trevizan – Active-Learning for Learned Search Guidance Improvement (bibtex)
  • Arnaud Lequen – Learning Interpretable Classifiers for PDDL Planning (pdf, bibtex) (ECAI 2024 paper)
  • Mariam Musavi, Johannes K. Fichte, and Daniel Gnad – Learning Object Priorities for Partial Grounding (pdf, bibtex)
  • Windy Phung, Arnaud Lequen, Dominik Drexler, and Jendrik Seipp – LLM-Evolved Pattern Generators for Optimal Classical Planning (pdf, bibtex)

15:30 - 15:50 Coffee Break

15:50 - 17:30 Session: Open Discussion

  • Open Discussion
  • Closing Remarks

Topics of Interest

As per the above, the workshop includes any topic that falls into the following problem space, roughly classified along three dimensions:

  1. Data-driven artifacts: Learned or ML-generated planning and scheduling models (e.g., LLM-generated PDDL, or learned transition probabilities and environment predictions); learned action-decisions (e.g., action policies, components thereof and previous plans); learned search guidance (e.g., heuristics and state rankings); and combinations thereof.
  2. Objectives: Reliability in whatever form, including risk, safety, robustness, fairness, error bounds, etc., alongside possibly other concerns such as scalability and data efficiency, system design/engineering principles and challenges, and the interactions of these with reliability.
  3. Methodologies: Planning and scheduling algorithms in the presence of learned artifacts as per (1); analyzing such learned artifacts (quality assurance, reasoning, verification, testing, etc.); making such analyses amenable to human users (e.g., visualization, interaction); and potentially others as relevant to the objectives as per (2).

Some example points in this problem space are:

  • Safe reinforcement learning, methods that guarantee actions remain within safety limits during learning and/or execution.
  • Safeguarding of learned action policies through techniques such as monitoring, shielding, lookahead search, planning as a safety guardrail, temporal-logic constraints, barrier functions.
  • Quality assurance for LLM-generated planning models.
  • Safeguarding and quality assurance for LLM-based planning, e.g., reliability of chain-of-thought approaches and LLM-generated plans.
  • Reliability of learned planning models, like (structured) action and environment models incorporating data-driven predictions, e.g., in the face of sparse, noisy, and/or out-of-distribution data.
  • Data-driven model refinement.
  • Verifying or testing safety, robustness, goal-reaching guarantees, or other desirable properties of learned action policies and planning-control knowledge. Irreversible actions/no free exploration: settings where trial-and-error is fundamentally infeasible because a single failure can cause unacceptable harm, and high-fidelity simulation may be impractical.
  • Conservative/risk-sensitive learning: optimizing safety-aware objectives (e.g., worst-case, CVaR) rather than maximizing expected return alone.
  • Offline-to-online transition & sim-to-real robustness: safely moving from offline data or simulation to real deployment without early-stage performance degradation or safety violations.
  • Interpretability & verifiability: ensuring learned behavior is explainable and amenable to auditing in deployment-critical contexts.
  • Capability awareness/uncertainty estimation: enabling agents to recognize distributional shift or uncertainty and respond conservatively or defer appropriately, adapting to non-stationary environments.
  • Diagnosis of systems involving ML components.
  • Risk analysis of planning and scheduling with data-driven models.
  • Addressing the optimizer’s curse (the tendency of an optimizer to find extrapolation errors in learned models).
  • Bias in data-driven models.
  • Interactive visualizations enabling users to understand a planning/scheduling model or a learned action policy.

Important Dates

  • Paper submission deadline: May 15, 2026 (AoE) Extended deadline: May 22, 2026 (AoE).
  • Paper acceptance notification: June 9, 2026

Submission Details

All papers must be formatted like at the main conference (ICAPS author kit). Submitted papers should be anonymous for double-blind reviewing. Paper submission is via EasyChair.

We call for two kinds of submissions:

  • Technical papers, of length up to 8 pages plus unlimited references and appendices. The workshop is meant to be an open and inclusive forum, and we encourage papers that report on work in progress.
  • Position papers, of length up to 4 pages plus unlimited references and appendices. Given that reliability of data-driven planning and scheduling is rather new at ICAPS, we encourage authors to submit positions on what they believe are important challenges, questions to be considered, approaches that may be promising. We will include any position relevant to discussing the workshop topic. We expect to group position paper presentations into a dedicated session, followed by an open discussion.

Every submission will be reviewed by members of the program committee according to the usual criteria such as relevance to the workshop, significance of the contribution, and technical quality.

Policy on Previously Published Materials

Please do not submit papers that are already accepted for the ICAPS main conference. All other submissions are welcome. Authors submitting papers rejected from the ICAPS main conference, please ensure you do your utmost to address the comments given by ICAPS reviewers. Also, it is your responsibility to ensure that other venues your work is submitted to allow for papers to be already published in “informal” ways (e.g., on proceedings or websites without associated ISSN/ISBN).

Committee

  • Daniel Höller, Saarland University, Germany
  • Nitay Alon, Hebrew University of Jerusalem, Israel
  • Guy Azran, Technion – Israel Institute of Technology, Israel
  • Sarah Eisenstein-Keren, Technion – Israel Institute of Technology, Israel
  • Timo P. Gros, German Research Center for Artificial Intelligence, Germany
  • Jörg Hoffmann, Saarland University, Germany
  • Sarath Sreedharan, Colorado State University, USA
  • Marcel Steinmetz, French National Centre for Scientific Research (CNRS), France
  • Sylvie Thiebaux, University of Toulouse, France, and Australian National University, Australia
  • Felipe Trevizan, Australian National University, Australia
  • Marcel Vinzent, Saarland University, Germany
  • Eyal Weiss, Bar-Ilan University, Israel