Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autonomous Task Completion Based on Goal-directed Answer Set Programming

Alexis R. Tudor

TL;DR

This paper addresses the challenge of explainable autonomous task planning in complex environments, arguing that inductive logic programming and answer set programming offer interpretable reasoning that complements deep learning. It proposes a goal-directed ASP approach implemented via the s(CASP) system, integrated with a Python harness to interact with the VirtualHome simulator and generate executable plans. Key contributions include a Python-based harness to translate simulator state into s(CASP) facts, a set of planning rules (e.g., initial_state, transform, choose_action, update, complete_task), and a dynamic dependency-graph optimization that substantially reduces computation time. The work also identifies open issues such as scalable real-world state representation, time modeling, and avenues for combining ILP/ASP with larger-scale learning, outlining a path toward more reliable, auditable task execution in autonomous agents.

Abstract

Task planning for autonomous agents has typically been done using deep learning models and simulation-based reinforcement learning. This research proposes combining inductive learning techniques with goal-directed answer set programming to increase the explainability and reliability of systems for task breakdown and completion. Preliminary research has led to the creation of a Python harness that utilizes s(CASP) to solve task problems in a computationally efficient way. Although this research is in the early stages, we are exploring solutions to complex problems in simulated task completion.

Autonomous Task Completion Based on Goal-directed Answer Set Programming

TL;DR

This paper addresses the challenge of explainable autonomous task planning in complex environments, arguing that inductive logic programming and answer set programming offer interpretable reasoning that complements deep learning. It proposes a goal-directed ASP approach implemented via the s(CASP) system, integrated with a Python harness to interact with the VirtualHome simulator and generate executable plans. Key contributions include a Python-based harness to translate simulator state into s(CASP) facts, a set of planning rules (e.g., initial_state, transform, choose_action, update, complete_task), and a dynamic dependency-graph optimization that substantially reduces computation time. The work also identifies open issues such as scalable real-world state representation, time modeling, and avenues for combining ILP/ASP with larger-scale learning, outlining a path toward more reliable, auditable task execution in autonomous agents.

Abstract

Task planning for autonomous agents has typically been done using deep learning models and simulation-based reinforcement learning. This research proposes combining inductive learning techniques with goal-directed answer set programming to increase the explainability and reliability of systems for task breakdown and completion. Preliminary research has led to the creation of a Python harness that utilizes s(CASP) to solve task problems in a computationally efficient way. Although this research is in the early stages, we are exploring solutions to complex problems in simulated task completion.

Paper Structure

This paper contains 10 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background and Relevant Literature
  3. Methodology
  4. Status
  5. Preliminary Results
  6. Open Issues and Expected Achievements
  7. Representing a Complex Real-World State
  8. The Passage of Time
  9. Large-scale Learning
  10. Conclusion

Figures (3)

  • Figure 1: An image showcasing the VirtualHome simulator and an example of its associated task instructions puig2018.
  • Figure 2: A diagram showing the high-level functionality of the Python harness for s(CASP). The Python harness can perform actions in the VirtualHome environment, and then convert the state of the environment to s(CASP) facts. These facts can then be used to inform the next action of the agent.
  • Figure 3: An example dependency graph for a family tree program where niece is the queried rule.