GLaMP: A Grounded Language Model-Based Multi-Agent System for Long-Horizon Robotic Task Planning in Industrial Settings
HONGPENG CHEN, David Navarro-Alarcon, PAI ZHENG
AI summary
Problem
Long-horizon industrial robotic planning struggles with predicate drift, symbolic-physical misalignment, and poor error recovery in open-loop pipelines. Existing methods lack reliable mechanisms to maintain consistency across complex, multi-step assembly tasks.
Approach
GLaMP uses a vision-language model to extract hierarchical task graphs from manuals, a perception agent to continuously ground real-time observations into PDDL predicates, and an LLM to compile reactive behavior trees with tiered recovery strategies for online error correction.
Key results
- Achieves 82–92% success rates across five industrial tasks
- Outperforms baselines by 5–20% in success rate with lower latency
- Ablation confirms task graphs, reactive trees, and upstream feedback are essential
- Enables typed failure classification and adaptive parameter regeneration to prevent predicate drift
Why it matters
Provides a reliable, interpretable planning paradigm for industrial automation, enabling language-driven robots to execute complex, long-horizon assembly tasks with minimal human intervention.
Abstract
This paper proposes GLaMP, a grounded language model–based multi-agent system for long-horizon robotic task planning in industrial settings. GLaMP uses a vision- language model (VLM) agent to infer a hierarchical task graph from manuals, grounds multimodal observations into planning domain definition language (PDDL) predicates to maintain symbolic consistency, and employs an large language model (LLM) behavior-tree planner for interpretable plan generation and execution. Typed-symbolic feedback enables failure-aware re-grounding and fallback replanning. Experiments on five industrial tasks show that GLaMP consistently outperforms representative baselines in success rate, planning latency, and execution time.