Interested in learning more? Check out our project page below!

Project Page: plan-with-climb.github.io
Paper: arxiv.org/pdf/2410.13756
[6/6]

LLMs by themselves are not perfect zero-shot solvers and can sometimes misinterpret or insufficiently describe the dynamics of an environment. CLIMB empowers these systems to explore, identify, and correct these biases to solve complex tasks.

[5/6]

We develop BlocksWorld++, a curriculum of block stacking and manipulation tasks that showcases CLIMB’s ability to generalize and reuse task primitives. We implemented BlocksWorld++ in logical PDDL, IsaacSim, and a real tabletop environment to enable evaluations at multiple levels of fidelity.

[4/6]

When given multiple tasks in the same environment, CLIMB caches and reuses the model generated on previous tasks when solving new ones, adding capabilities with each task completed. Given a diverse curriculum CLIMB can solve tasks more efficiently by leveraging knowledge it gained previously.

[3/6]

With CLIMB, all the user needs to provide is a description of the environment and a list of tasks to accomplish. CLIMB builds an estimated world model of its domain in PDDL, calls a symbolic planner to determine an initial task plan, and attempts to solve the task.

[2/6]