Instructing Robots with Natural Language via Bi-RNNs for Temporal Logic Translation

Published

4-12-2023

Type of Work

Article

Abstract

We consider the problem of planning trajectories that satisfy natural language instruction. We explore translating natural language commands to temporal logic formulae to resolve ambiguities for planning. Our main contribution is a new bi-directional recurrent neural network (Bi-RNN) architecture for this translation task. We experimentally show that the proposed Bi-RNN architecture achieves 1.6% better accuracy, 20% faster inference time, and 98% faster training time compared to leading models owing to bidirectional processing. The overall system, including a planning algorithm, exhibits useful diverse behaviours that satisfy given instructions.

Citation

Sharma, S., Lee, K. M. B., Brown, M., & Best, G. (2023). Instructing robots with natural language via Bi-RNNs for temporal logic translation. In Australasian Conference on Robotics and Automation. https://ssl.linklings.net/conferences/acra/acra2023_proceedings/views/includes/files/pap113s2.pdf

Link to Full Text

Share

COinS