For more details on our work, please see
Paper: arxiv.org/abs/2511.01107
Open-source code: github.com/isabelliu0/S...

🙌 This work was done with my incredible advisors @tomssilver.bsky.social and @ben-eysenbach.bsky.social and amazing collaborator Bowen Li.

In Cleanup Table environment:
(8/8)

In Cluttered Drawer environment:
(7/8)

We demonstrate our results on PyBullet simulated environments. In Obstacle Tower environment:
(6/8)

3️⃣ Self-Balancing. SLAP automatically navigates this spectrum between planning and learning. In the extremes, if shortcuts are too difficult to learn, SLAP reduces to the original planners; if the tasks are easy, SLAP reduces to RL—the plan collapses into a single shortcut.
(5/8)

2️⃣ Hierarchical. SLAP leverages the hierarchical structure of planners to decompose the unstructured, long-horizon problem into smaller, tractable shortcuts for RL to learn.
(4/8)

Our approach is:
1️⃣ Automatic. SLAP does not require any intermediate inputs or additional assumptions. The way we find, learn, and deploy shortcuts is completely automatic.
(3/8)

We introduce Shortcut Learning for Abstract Planning (SLAP), a new method that uses reinforcement learning (RL) to discover shortcuts in the planning graphs induced by task and motion planning (TAMP) skill libraries. It is a plug-and-play module that can be trained on top of existing planners.
(2/8)