From Reals to Logic and Back: Inventing Symbolic Vocabularies, Actions and Models for Planning from Raw Data

Traditional robot planning relies on human-crafted logic representations, but this paper introduces a method to autonomously learn abstract representations from raw robot data. Results show these learned models enable scalable planning for complex tasks without human intervention.

Hierarchical Planning and Learning for Robots in Stochastic Settings Using Zero-Shot Option Invention

This paper proposes a new method for robots to plan actions in complex environments, even when the environment is unknown. The robot learns to create its own high-level actions without needing pre-programmed ones. This allows the robot to quickly solve new problems in unseen environments. The method is shown to be faster and achieve significantly better solutions than existing approaches

Using Deep Learning to Bootstrap Abstractions for Robot Planning

In this paper, we use deep learning to identify critical regions and automatically construct hierarchical state and action abstractions. We use these hierarchical abstractions with a multi-source mutli-directional hierarchical planner to compute solutions for robot planning problem.

Learning and Using abstractions for Robot Planning

In this paper, we propose unified framework based on deep learning that learns sound abstractiosn for complex robot planning problems and uses it to efficiently perform hierarchical planning.

Anytime Task and Motion Policies for Stochastic Envrionments

In this paper, we provide and efficient abstraction based methods to compute task and motion policies for complex robotics task for stochastic environments.