VernaCopter: Disambiguated Natural-Language-Driven Robot via Formal Specifications

TL;DR

VernaCopter addresses NL-driven robot control by introducing a formal-spec bridging approach using Signal Temporal Logic ($STL$). The system uses a planning assistant LLM to generate $STL$ specifications, verified by syntax and semantics checkers, and solved by a model-checking optimizer; this reduces NL ambiguity and LLM uncertainty compared to pure NL prompts. Experimental results on reach-and-avoid and treasure-hunt tasks show 100% success and collision-free performance for VernaCopter versus substantially lower rates for conventional NL-based planners. The work demonstrates that formal specifications can enable reliable NL-driven robotics and suggests future directions in decomposition and LLM fine-tuning.

Abstract

It has been an ambition of many to control a robot for a complex task using natural language (NL). The rise of large language models (LLMs) makes it closer to coming true. However, an LLM-powered system still suffers from the ambiguity inherent in an NL and the uncertainty brought up by LLMs. This paper proposes a novel LLM-based robot motion planner, named \textit{VernaCopter}, with signal temporal logic (STL) specifications serving as a bridge between NL commands and specific task objectives. The rigorous and abstract nature of formal specifications allows the planner to generate high-quality and highly consistent paths to guide the motion control of a robot. Compared to a conventional NL-prompting-based planner, the proposed VernaCopter planner is more stable and reliable due to less ambiguous uncertainty. Its efficacy and advantage have been validated by two small but challenging experimental scenarios, implying its potential in designing NL-driven robots.

Figures (7)

• Figure 1: A treasure-hunt quadcopter with interdependent subtasks.
• Figure 2: Facilitating LLM-based robots with STL specifications.
• Figure 3: The VernaCopter planner architecture.
• Figure 4: The 2D views of the two experimental scenarios.
• Figure 5: Generated paths for R&A.