ICAPS 2026 Summer School

Classical planning studies how an intelligent agent can automatically compute a sequence of actions that achieves a desired goal. Despite its simplifying assumptions of deterministic actions and full observability, classical planning lies at the core of automated planning research and has provided many of the concepts and algorithms that continue to shape research across automated planning and decision making. In this tutorial, I will introduce the main approaches to solving classical planning problems, including explicit state-space search, symbolic search, and SAT-based planning. I will then focus on heuristic search, the dominant paradigm in domain-independent planning, and present some of its most influential ideas, such as delete-relaxation and abstraction-based heuristics.

For over six decades, automated planning has been central to AI, enabling systems to reason about actions and achieve goals in complex, dynamic environments. As large language models (LLMs) redefine what AI systems can do, planning ideas are becoming newly relevant: they offer principled ways to specify objectives and constraints, evaluate behavior, and move from fluent language to reliable, goal-directed decision making.

This lecture focuses on methods for planning with language models. The first part covers plan generation with LLMs, treating the model as a planner. We will cover both post-training and inference. The second part explores hybrid approaches that combine LLMs with symbolic planning components at build time, inference time, or both, to improve correctness, scalability, and robustness.

For over six decades, automated planning has been central to AI, enabling systems to reason about actions and achieve goals in complex, dynamic environments. As large language models (LLMs) redefine what AI systems can do, planning ideas are becoming newly relevant: they offer principled ways to specify objectives and constraints, evaluate behavior, and move from fluent language to reliable, goal-directed decision making.

This lecture focuses on methods for planning with language models. The first part covers plan generation with LLMs, treating the model as a planner. We will cover both post-training and inference. The second part explores hybrid approaches that combine LLMs with symbolic planning components at build time, inference time, or both, to improve correctness, scalability, and robustness.

The Planning Domain Definition Language (PDDL) provides a compact and expressive way to model the transition systems underlying classical planning problems, while explicitly exposing structural properties that modern planners can exploit to scale to large combinatorial search spaces.

This lab will be a hands-on session introducing the core building blocks of PDDL. Participants will progressively model planning problems of increasing complexity, gaining practical experience with common modeling patterns and understanding how different modeling choices can significantly affect planner performance and scalability.

Participants will become familiar with state-of-the-art planners deployed through planning.domains, use the online environment editor.planning.domains, and work with an offline development setup based on Visual Studio Code and a PDDL plugin. By the end of the lab, attendees will be equipped with practical skills to confidently model and debug planning models.

Generalized planning is about solving entire collections of planning problems over a common domain. That is, computing a general policy that solves every instance in the collection. For this, abstractions for states, transitions and even paths must be computed and aligned in sound and effective ways.

In this lecture, we focus on learning approach for generalized planning. In the first part, we focus on the task of learning policies from training sets composed of small instances, in the symbolic and neural settings. In the second part, we focus on the task of learning the planning models required in generalized planning under different set of assumptions.

This course provides a comprehensive introduction to probabilistic planning -- the art of making optimal sequential decisions when actions have uncertain outcomes. Starting from the familiar ground of classical AI planning, we build up the mathematical foundations of Markov Decision Processes (MDPs) and Stochastic Shortest Path (SSP) problems. We then explore various algorithms ranging from dynamic programming (Value Iteration, Policy Iteration) through heuristic search (LAO*, LRTDP) to Monte Carlo simulation methods (UCT, POMCP, GoalHSVI). We also venture beyond classical SSPs to address real-world challenges such as non-avoidable dead-end states, safety constraints expressed in temporal logic (PCTL), and the modeling languages -- PPDDL and RDDL -- that have driven two decades of progress through the International Probabilistic Planning Competition. The course closes with simulator-based planning using the scikit-decide framework and an application to probabilistic flight planning under wind uncertainty.

Theory and practice are tightly interleaved throughout the 3-hour session. We illustrate each algorithmic concept with worked examples on small MDPs, then bring it to life through demonstrations of running solvers on a stochastic maze directly from the slides, observing in real time how different algorithms explore the state space and react to uncertainty. These live visualizations make abstract notions like Bellman backups, greedy envelopes, and Monte Carlo rollouts tangible. Furthermore, the progression from foundational models to competition benchmarks and industrial applications gives attendees both the theoretical grounding and the practical intuition needed to apply probabilistic planning to their own research problems.

Many decision problems come with rich, structured knowledge of dynamics and rewards -- so why discard that structure and go model-free? This lab makes the case for model-based probabilistic planning and walks through the full pipeline: formal modeling, an interactive environment, and a differentiable solver. We’ll start with a concise, pragmatic refresher on MDPs and stochastic planning, positioning RDDL relative to PPDDL and highlighting support for concurrency, numeric/Boolean fluents, conditional effects, and exogenous processes. We then introduce pyRDDLGym, which compiles RDDL descriptions into OpenAI Gym–compatible environments for fast simulation, visualization, and benchmarking.

The centerpiece is a hands-on Colab session where participants will:
(i) model a stochastic control task in RDDL;
(ii) generate and explore a pyRDDLGym environment; and
(iii) experiment with the default random agent and a planning-by-backpropagation solver—optimizing a parameterized policy by differentiating through model rollouts, bridging planning and learning.

By the end, participants will be able to model and solve stochastic planning problems using RDDL and pyRDDLGym, and gain understanding when explicit models can outperform black-box methods.

Scheduling is about deciding when and with what resources set of tasks should be performed. Unlike planning, we typically know what tasks are to be performed, or perhaps must select them from some options, or the tasks are given to us over time in an online setting. However, we do not typically have to decide what actions to perform. Scheduling is widely studied in Operations Research (OR), theoretical computer science, computer systems, and queueing theory, as well as in AI, leading to a very large variety of problems and solution approaches. In this tutorial, I will provide a brief overview of this scope and then focus on exact techniques for problems studied in OR and AI: mixed-integer linear programming, constraint programming, and recent work on using heuristic search to solve scheduling problems within the domain-independent dynamic programming framework.

You’re a new robot, just coming online. How do you move your body? What’s surrounding you? What are you supposed to do? Come to this session to find out!

In reinforcement learning (RL) a physical or virtual agent needs to learn about its surroundings while also figuring out how to maximize a reward. These sessions will help you understand
1) how to identify when RL could be the right framing for a problem,
2) what makes an RL problem more or less difficult to solve,
3) how a handful of simple RL algorithms work, and
4) where to go if you’d like to learn more.

In this lab, we explore the implementation and performance of both tabular and Deep Reinforcement Learning (RL) algorithms.

First, we implement two foundational tabular RL algorithms, Q-learning and SARSA, within a grid world environment to analyze how off-policy and on-policy methods converge on different optimal policies. We then conduct an hyperparameter tuning exercise to evaluate how key variables, such as the learning rate and exploration rate, influence agent performance.

Finally, we transition to the Deep RL setting by implementing aspects of the Deep Q-Network (DQN) algorithm, where we similarly investigate the impact of hyperparameter choices on convergence.

Colab Notebook

Automated planning is especially challenging when state and action spaces are continuous, time horizons are long, environments are constrained, and feedback is sparse---all of which are common in robotics. Task and motion planning (TAMP) addresses these challenges by enabling the agent to reason jointly about “what to do” (task planning) and “how to do it” (motion planning). These two levels of reasoning are often entangled due to the lossy nature of the task-planning abstractions that translate the continuous robot environments into discrete representations. In this tutorial and lab, we will approach these challenges with a first-principles introduction to TAMP. Participants should come away with new intuitions and practical tools---an understanding of both “what to do” and “how to do it” when decision-making calls for TAMP.

Automated planning is especially challenging when state and action spaces are continuous, time horizons are long, environments are constrained, and feedback is sparse---all of which are common in robotics. Task and motion planning (TAMP) addresses these challenges by enabling the agent to reason jointly about “what to do” (task planning) and “how to do it” (motion planning). These two levels of reasoning are often entangled due to the lossy nature of the task-planning abstractions that translate the continuous robot environments into discrete representations. In this tutorial and lab, we will approach these challenges with a first-principles introduction to TAMP. Participants should come away with new intuitions and practical tools---an understanding of both “what to do” and “how to do it” when decision-making calls for TAMP.

Deploying robots in real-world environments requires both the ability to reason about long-horizon goals and the ability to acquire knowledge from experience. Classical planning offers principled frameworks for structured reasoning, goal decomposition, and guaranteed-property search but assumes models that are rarely available in practice and is often prohibitively expensive at inference time for robotics applications. Machine learning offers powerful tools for acquiring those models, discovering useful representations, and adapting to novel situations, but struggles with sample efficiency, safety, and compositional generalization without structure. How can we obtain the best of both worlds? In this tutorial, we will explore different ways of combining learning and planning through the use of abstractions, learned representations, and learned models.