[Paper Review] Axiom Learning and Belief Tracing for Transparent Decision Making in Robotics

Tiago Mota, Mohan Sridharan|arXiv (Cornell University)|Oct 20, 2020

AI-based Problem Solving and Planning8 references4 citations

TL;DR

This paper proposes an integrated robotics architecture that combines non-monotonic logical reasoning, deep learning, and decision-tree induction to enable robots to generate on-demand, relational explanations of decisions, beliefs, and hypothetical actions. By learning new axioms from data and using them to guide reasoning, the system significantly improves explanation accuracy—achieving 100% precision and recall in plan and belief justification tasks when learned axioms are used.

ABSTRACT

A robot's ability to provide descriptions of its decisions and beliefs promotes effective collaboration with humans. Providing such transparency is particularly challenging in integrated robot systems that include knowledge-based reasoning methods and data-driven learning algorithms. Towards addressing this challenge, our architecture couples the complementary strengths of non-monotonic logical reasoning, deep learning, and decision-tree induction. During reasoning and learning, the architecture enables a robot to provide on-demand relational descriptions of its decisions, beliefs, and the outcomes of hypothetical actions. These capabilities are grounded and evaluated in the context of scene understanding tasks and planning tasks performed using simulated images and images from a physical robot manipulating tabletop objects.

Motivation & Objective

• To enable robots to provide on-demand, relational explanations of decisions, beliefs, and hypothetical actions in complex, dynamic environments.
• To address the challenge of explainability in integrated systems combining knowledge-based reasoning and data-driven learning.
• To develop a system that learns new axioms governing action preconditions and effects from experience, improving transparency.
• To enable automatic tracing of belief evolution from initial states through sequences of axioms.
• To evaluate the architecture in both simulated and physical robot scenarios involving scene understanding and planning.

Proposed method

• The architecture uses Answer Set Programming (ASP) for non-monotonic logical reasoning to model domain dynamics and support belief tracing.
• Deep learning models are adapted for perception tasks such as object and spatial relation recognition from images.
• Decision-tree induction is used to learn new axioms (preconditions and effects) from observed data and reasoning outcomes.
• Explanations are constructed by retrieving relevant literals from reasoning traces, using both known and learned axioms.
• The system dynamically triggers learning when existing knowledge is insufficient to justify decisions or beliefs.
• Explanations are generated as relational descriptions of objects, actions, and attributes, grounded in logical inference and learned constraints.

Experimental results

Research questions

• RQ1 How can a robot system generate accurate, on-demand explanations of its decisions and beliefs using a combination of knowledge-based reasoning and data-driven learning?
• RQ2 To what extent does learning new axioms improve the precision and recall of explanation generation in robotics tasks?
• RQ3 Can the system automatically trace the evolution of a belief from initial knowledge to current state using logical inference?
• RQ4 How does the integration of reasoning and learning enhance the transparency and reliability of robot decision-making?
• RQ5 What is the impact of learned axioms on explaining hypothetical or counterfactual scenarios in planning and perception?

Key findings

• When learned axioms were used for reasoning, precision and recall for explanation generation reached 100% in plan description and belief justification tasks in simulated scenes.
• The use of learned axioms significantly improved performance, especially for questions about unexecuted actions, where recall increased from 63.01% to 100% in physical robot experiments.
• Even without learned axioms, the system maintained reasonable performance (e.g., 71.85% precision in plan description), indicating robustness from existing knowledge.
• Errors in explanations were rare and typically involved over-specification (extra literals), suggesting high fidelity in explanation construction.
• Performance dropped substantially when axioms related to the goal were removed, confirming their critical role in accurate explanation generation.
• The system demonstrated robustness to sensor and actuation noise, with recognition, constraint learning, and manipulation errors at 15%, 5–10%, and 15% respectively, while still producing accurate explanations.