4th Workshop on

Learning Effective Abstractions for Planning (LEAP)

CoRL 2026  ·  Nov 12, 2026  ·  Austin, TX, USA

Email: leap-workshop@googlegroups.com

Overview

Complex long-horizon and sparse reward robotics tasks are challenging for scaling end-to-end learning methods. By contrast, planning approaches have shown great potential to handle such complex tasks effectively. One of the major criticisms of planning-based approaches has been the lack of availability of accurate world models (aka abstractions) to utilize.

There has been a renewed interest in using learning-based approaches to learn symbolic representations that support planning. However, this research is often fragmented into disjoint sub-communities such as task and motion planning, reinforcement learning (hierarchical, model-based), planning with formal logic, planning with natural language (language models), and neuro-symbolic AI. As in previous years, LEAP will bring together researchers from disparate subfields to discuss how we can combine learning, planning, and abstractions to solve increasingly complex long-horizon and sparse-reward robotics tasks.

As in prior editions, the workshop continues to engage with the foundational questions that have defined LEAP since its inception:

  1. What is the right objective for abstraction learning for robotic planning? To what extent should we account for soundness, completeness, the target planner and its planning efficiency, and the task distribution?
  2. What level of abstraction is necessary for effective robot decision-making? How general-purpose or specific must these abstractions be for long-term autonomy, and do learned abstractions need to be hierarchical?
  3. When, where, and from what data should abstractions be learned? Should they be priors built in the robot factory, derived from expert demonstrations, or discovered "in the wild" through interaction between the robot and its environment? What are the trade-offs between top-down operator construction and bottom-up operator discovery, and how do they impact planning efficiency and generalization?

Focal Themes of LEAP 2026

While the questions above ground the workshop, our key attention this year is deliberately focused on two specific, concrete, and timely challenges in learning abstractions for planning.

💻 Theme #1: Code as Abstractions

Programs and code offer a uniquely powerful substrate for abstraction in robot planning. They are compositional, verifiable, reusable, and directly executable, and they interface naturally with modern foundation models that can read, write, and edit code. We are eager to explore and discuss when and how code is the right abstraction for robot decision-making, as well as to identify its limitations:

  • When is code (viz., programs, policies-as-code, reward and constraint specifications, and skill libraries) the right abstraction for robot planning? When does it break down? What soundness, completeness, and verifiability guarantees can code-based abstractions provide that other representations cannot?
  • How can pre-trained foundation models (LLMs, VLMs, VLAs) be used to write, refine, and repair code-based abstractions? How can we incorporate cost-sensitive reasoning, plan quality, and executability such that generated programs produce efficient, safe plans for embodied agents?
  • How should code-based abstractions be grounded in perception and action, learned and reused across tasks, and combined with bottom-up skill discovery, so that they facilitate imitation and reinforcement learning and remain reliable in open-world settings?

🌐 Theme #2: World Models as Abstractions

Learned world models promise abstractions of environment dynamics that can be searched, simulated, and planned over. We want to examine what makes a world model a good abstraction for planning (rather than merely for prediction), and how planning can be performed in, on top of, or adjacent to such models:

  • What makes a world model a good abstraction for planning rather than merely for prediction? At what level of abstraction (state, latent, object-centric, symbolic) should world models operate, and do they need to be hierarchical?
  • How can we plan within, on top of, or adjacent to learned world models? What are the trade-offs between planning directly within a learned model, distilling it into a planner-friendly abstraction, and using it to guide reinforcement or imitation learning?
  • When, where, and from what data should world-model abstractions be learned (priors built in the factory, expert demonstrations, or "in the wild" interaction), and how do we ensure the resulting plans are safe, cost-effective, and robust to model error?

Areas of Interest

We solicit papers on the following topics:

Submission Guidelines

We solicit workshop paper submissions relevant to the above call of the following types:

  • Long papers — up to 8 pages plus unlimited references / appendices
  • Short papers — up to 4 pages plus unlimited references / appendices

Please format submissions in CoRL or IEEE conference (ICRA or IROS) styles. Submissions do not need to be anonymized. To authors submitting papers rejected from other conferences: please ensure that comments given by the reviewers are addressed prior to submission.

Note: Please feel free to submit work under review or accepted for presentation at other workshops and/or conferences as we will not require copyright transfer.

📄 Submit on OpenReview →
Note: The CoRL workshop organizers have requested that we do not accept submissions that are already accepted at the main CoRL conference. We kindly ask authors to respect this policy when submitting to our workshop.

Important Dates

Format & Schedule

  1. 1
    A small, focused set of invited talks. We cap the number of invited speakers at three, giving each ample time for a substantive, thought-provoking talk. This year, we have asked our invited speakers to foreground the open questions in abstraction learning for robot planning, rather than recapping results, so that their talks directly set up the breakout and panel discussions that follow.
  2. 2
    Spotlight and lightning talks for contributed work. Every accepted paper receives a 2-minute lightning talk, and a selection of standout contributions receive a 5-minute spotlight, keeping the contributed program visible and energetic while protecting time for discussion.

The detailed schedule for LEAP 2026 will be announced closer to the workshop date.

Invited Speakers

Tomás Lozano-Pérez
Tomás Lozano-Pérez
MIT · USA
Dieter Fox
Dieter Fox
Ai2 & UW · USA
Fabio Ramos
Fabio Ramos
Nvidia & U. Sydney · USA/AUS
Panpan Cai
Panpan Cai
Shanghai Jiao Tong U. · China

Organizing Committees