Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neuro-Symbolic Generation of Explanations for Robot Policies with Weighted Signal Temporal Logic

Mikihisa Yuasa, Ramavarapu S. Sreenivas, Huy T. Tran

TL;DR

A neuro-symbolic explanation framework that generates a weighted signal temporal logic specification which describes a robot policy in a human-interpretable form which outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing accuracy.

Abstract

Neural network-based policies have demonstrated success in many robotic applications, but often lack human-explanability, which poses challenges in safety-critical deployments. To address this, we propose a neuro-symbolic explanation framework that generates a weighted signal temporal logic (wSTL) specification to describe a robot policy in a interpretable form. Existing methods typically produce explanations that are verbose and inconsistent, which hinders explainability, and loose, which do not give meaningful insights into the underlying policy. We address these issues by introducing a simplification process consisting of predicate filtering, regularization, and iterative pruning. We also introduce three novel explainability evaluation metrics -- conciseness, consistency, and strictness -- to assess explanation quality beyond conventional classification metrics. Our method is validated in three simulated robotic environments, where it outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing classification accuracy. This work bridges policy learning with formal methods, contributing to safer and more transparent decision-making in robotics.

Neuro-Symbolic Generation of Explanations for Robot Policies with Weighted Signal Temporal Logic

TL;DR

A neuro-symbolic explanation framework that generates a weighted signal temporal logic specification which describes a robot policy in a human-interpretable form which outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing accuracy.

Abstract

Neural network-based policies have demonstrated success in many robotic applications, but often lack human-explanability, which poses challenges in safety-critical deployments. To address this, we propose a neuro-symbolic explanation framework that generates a weighted signal temporal logic (wSTL) specification to describe a robot policy in a interpretable form. Existing methods typically produce explanations that are verbose and inconsistent, which hinders explainability, and loose, which do not give meaningful insights into the underlying policy. We address these issues by introducing a simplification process consisting of predicate filtering, regularization, and iterative pruning. We also introduce three novel explainability evaluation metrics -- conciseness, consistency, and strictness -- to assess explanation quality beyond conventional classification metrics. Our method is validated in three simulated robotic environments, where it outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing classification accuracy. This work bridges policy learning with formal methods, contributing to safer and more transparent decision-making in robotics.
Paper Structure (28 sections, 1 theorem, 15 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 28 sections, 1 theorem, 15 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Weighted Signal Temporal Logic
  4. Normalized wSTL
  5. Problem Statement
  6. Methodology
  7. Neural Network Architecture
  8. Simplification Process
  9. Predicate Filter
  10. Loss Regularization
  11. Weight Pruning
  12. Trajectory Sampling
  13. Explainability Evaluation Metrics
  14. Conciseness
  15. Consistency
  16. ...and 13 more sections

Key Result

Theorem 1

A normalized wSTL formula $\phi$ in CNF can be represented as follows, where $n, m_{i}\in\mathbb{Z}_{>0}$, $w_{i}, w_{ij}\in\mathbb{R}_{>0}$, $\tilde{w}_{ij}:=w_{i}w_{ij}$, $\sum_{i=1}^{n}w_{i}= \sum_{j=1}^{m_i}w_{ij} = 1$, and $\psi_{ij}$ are predicates. Furthermore, the weights $\tilde{w}_{ij}$ satisfy $\sum_{i=1}^{n}\sum_{j=1}^{m_i}\tilde{w}_{ij}= 1$.

Figures (4)

  • Figure 1: Existing neuro-symbolic methods often produce explanations that are verbose, inconsistent, and loose. Our method, TLNet, addresses this by generating concise, consistent, and strict wSTL specifications.
  • Figure 2: We propose a neuro-symbolic approach for generating an explanation of a given robot policy using wSTL. Our network architecture, TLNet, includes a simplification process that ensures resulting explanations balance classification accuracy with conciseness, consistency, and strictness.
  • Figure 3: Example TLNet architecture with two predicates, where $\phi_\mathbf{w}$ is the explanation generated from $\mathbf{w}$.
  • Figure 4: Initial states of our test environments with relevant objects shown.

Theorems & Definitions (2)

  • Theorem 1: Syntactic Weight Distributivity of Boolean Operators
  • proof