Intent-driven LLM ensemble planning for flexible multi-robot manipulation

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Complex multi-action manipulation in multi-robot systems requires sophisticated task planning.
• Bridging human intent with robot execution is a significant challenge in autonomous systems.

Purpose of the Study:

• To develop an intent-driven task planning system for heterogeneous multi-robot cells.
• To generate precedence-consistent robot-action sequences from natural-language human intent.

Main Methods:

• A pipeline integrating perception-to-text encoding, large language models (LLMs) for action sequence generation, an LLM-based verifier, and a deterministic consistency filter.
• Evaluation on a collaborative EV battery dismantling task for recycling.

Main Results:

• The ensemble-with-verification approach improved full-sequence correctness from 0.761 to 0.824 and next-object correctness from 0.866 to 0.894.
• Human participant experiments showed low user effort and reduced time to execution.

Conclusions:

• The developed system reliably maps operator intent to safe multi-robot plans.
• The ensemble-with-verification method enhances the performance and usability of robot task planning.