Scale-Plan: Scalable Language-Enabled Task Planning for Heterogeneous Multi-Robot Teams

Piyush Gupta; Sangjae Bae; Jiachen Li; David Isele

cs.ROMar 9, 2026

Scale-Plan: Scalable Language-Enabled Task Planning for Heterogeneous Multi-Robot Teams

Authors: Piyush Gupta, Sangjae Bae, Jiachen Li, David Isele

Organizations: Honda Research Institute, San Jose, CA 95134, USA. · Georgia Institute of Technology, Atlanta, GA 30332, USA.

Abstract

Long-horizon task planning for heterogeneous multi-robot systems is essential for deploying collaborative teams in real-world environments; yet, it remains challenging due to the large volume of perceptual information, much of which is irrelevant to task objectives and burdens planning. Traditional symbolic planners rely on manually constructed problem specifications, limiting scalability and adaptability, while recent large language model (LLM)-based approaches often suffer from hallucinations and weak grounding-i.e., poor alignment between generated plans and actual environmental objects and constraints-in object-rich settings. We present Scale-Plan, a scalable LLM-assisted framework that generates compact, task-relevant problem representations from natural language instructions. Given a PDDL domain specification, Scale-Plan constructs an action graph capturing domain structure and uses shallow LLM reasoning to guide a structured graph search that identifies a minimal subset of relevant actions and objects. By filtering irrelevant information prior to planning, Scale-Plan enables efficient decomposition, allocation, and long-horizon plan generation. We evaluate our approach on complex multi-agent tasks and introduce MAT2-THOR, a cleaned benchmark built on AI2-THOR for reliable evaluation of multi-robot planning systems. Scale-Plan outperforms pure LLM and hybrid LLM-PDDL baselines across all metrics, improving scalability and reliability. Project website: https://github.com/honda-research-institute/Scale_Plan

Figures & tables

Fig. 1 : Task-relevant environment filtering for scalable planning. Naive long-horizon planning considers all detected objects and robot capabilities, leading to large action spaces and planning errors (left). Scale-Plan extracts only task-relevant scene information and skills (right), significantly reducing combinatorial complexity and enabling efficient multi-robot plan synthesis.

Fig. 4 : Overview of the Scale-Plan framework. Scale-Plan identifies task-relevant environment information by searching an action graph constructed from the PDDL domain. The filtered information is then used by an LLM-based planning pipeline to perform structured task decomposition, task allocation, and plan integration. The resulting high-level plan is translated into an AI2-THOR–executable plan by Plan-to-Code and executed in the simulator.

Panel of Figure 3 — Fig. 5 : Edge generation in the action graph. A strict edge $a_{1}\rightarrow a_{2}$ is added if $PRE(a_{2})\subseteq EFF(a_{1})$ . A relaxed edge is added when $PRE(a_{2})\cap EFF(a_{1})\neq\emptyset$ and strict connectivity is absent, ensuring graph connectivity without over-densification.

Task Category	Task	MAT-THOR Ground Truth	MAT2-THOR Ground Truth
	Wash the lettuce and place lettuce on the Countertop	[{“name”: “CounterTop”, “contains”: [“Lettuce”], “state”: None}]	[{“name”: “CounterTop”, “contains”: [“Lettuce”], “state”: None}, {“name”: “Lettuce”, “contains”: [], “state”: “CLEANED”} ]
	Toast a slice of the breadloaf	[{“name”: “Bread”, “contains”: [], “state”: “COOKED”}]	[{“name”: “Bread”, “contains”: [], “state”: “COOKED”}, {“name”: “Toaster”, “contains”: [“Bread”], “state”: “ON”}, {“name”: “Bread”, “contains”: [], “state”: “SLICED”} ]
	Place the Bar of soap in the sink first and then place dishsponge in the sink	[ {“name”: “Sink”, “contains”: [“SoapBar”, “ScrubBrush”] , “state”: None} ]	[ {“name”: “Sink”,“contains”:[“SoapBar”,“DishSponge”], “state”: None} ]
Simple	Put two vegetables in the fridge parallely	[{“name”: “Fridge”, “contains”: [“Potato”, “Lettuce”], “state”: None}]	[ {“name”: “Fridge”, “contains”: [“Potato”,“Lettuce”, “Tomato”], “state”: None, “num_contains”: 2} ]
	Wash the lettuce and place lettuce, tomato on the Countertop	[{“name”: “CounterTop”, “contains”: [“Lettuce”], “state”: None}]	[ {“name”: “CounterTop”, “contains”: [“Lettuce”, “Tomato”], “state”: None}, {“name”: “Lettuce”, “contains”: [], “state”: “CLEANED”} ]
	Put a plate of slice lettuce into the microwave	[{“name”: “Microwave”, “contains”: [“Plate”], “state”: None}]	[{“name”: “Microwave”, “contains”: [“Plate”], “state”: None}, {“name”: “Plate”, “contains”: [“Lettuce”], “state”: None}, {“name”: “Lettuce”, “contains”: [], “state”: “SLICED”} ]

TABLE I : The MAT2-THOR dataset is derived by cleaning and refining the original MAT-THOR dataset. During this process, inconsistent, missing, and erroneous ground-truth conditions ( red ) in MAT-THOR were corrected ( blue ). Additionally, we introduce an optional parameter, num_contains , which specifies the minimum number of containment conditions that must be satisfied for a task to be considered complete.

Method	Simple No. Tasks	Simple TCR (%)	Simple GCR (%)	Simple ER (%)	Complex No. Tasks	Complex TCR (%)	Complex GCR (%)	Complex ER (%)	Vague No. Tasks	Vague TCR (%)	Vague GCR (%)	Vague ER (%)	Overall TCR (%)	Overall GCR (%)	Overall ER (%)
LLM as a Planner	25	44	56	79	17	18	40	69	7	57	57	88	37	51	77
LLM+P (PDDL based)	25	48	57	74	17	24	39	61	7	43	52	58	39	50	67
LaMMA-P (PDDL plan only)	25	36	56	79	17	18	37	66	7	57	64	77	33	51	74
LaMMA-P (LLM corrected)	25	76	84	91	17	24	56	77	7	43	50	85	53	69	85
Scale Plan (Ours)	25	92	95	98	17	59	75	92	7	71	71	82	78	85	94

TABLE II : Quantitative performance comparison on MAT2-THOR in AI2-THOR. Metrics include Task Completion Rate (TCR), Goal Condition Recall (GCR), and Executability Rate (ER) across simple, complex, and vague task categories.

Method	Simple No. Tasks	Simple TCR (%)	Simple GCR (%)	Simple ER (%)	Complex No. Tasks	Complex TCR (%)	Complex GCR (%)	Complex ER (%)	Vague No. Tasks	Vague TCR (%)	Vague GCR (%)	Vague ER (%)	Overall TCR (%)	Overall GCR (%)	Overall ER (%)
No-EF + Joint Task Planning (JTP)	25	80	85	96	17	41	65	90	7	71	71	89	65	76	93
No-EF	25	80	89	97	17	44	70	91	7	71	71	90	66	80	94
LLM-Based Shallow Filtering (LLM-SF)	25	88	94	99	17	41	70	90	7	57	64	85	67	81	94
ScalePlan (Full Model)	25	92	95	98	17	59	75	92	7	71	71	82	78	85	94

TABLE III : Ablation results comparing performance with and without task-relevant environment filtering. Filtering improves task completion and scalability while maintaining robustness.

Method	Simple PT (s)	Complex PT (s)	Vague PT (s)	Overall PT (s)
LLM as a Planner	10.8	15.7	13.1	12.83
LLM+P (PDDL based)	21.4	22.2	25.9	22.32
LaMMA-P (PDDL plan only)	52.5	62.2	64.6	57.59
LaMMA-P (LLM corrected)	57.0	67.6	83.0	64.39
Scale Plan (Ours)	53.5	72.1	69.7	62.27

TABLE IV : Planning Time (PT) comparison across methods, measuring total computation time including LLM reasoning and symbolic planning.

Panel of Figure 8 — Fig. 6 : Representative failure cases observed in Scale-Plan. (a) Missing affordance reasoning leads to failure to open a receptacle before placing an object. (b) Missing manipulation constraints cause failure to release a held object before picking up another.

Explore similar work

Hierarchical LLM-Based Multi-Agent Framework with Prompt Optimization for Multi-Robot Task Planning
Feb 25, 2026Tomoya Kawabe, Rin TakanoLarge Language Model-Based Robot Planning Multi-Agent LLM Systems
Data Science Laboratories, NEC Corporation, 1753, Shimonumabe, Nakahara-ku, Kawasaki, Kanagawa, 211-8666, Japan.
HEART: Coordination of Heterogeneous Expert Agents for Physically Grounded Robotic Task Planning
Jun 24, 2026Junho Lee, Seabin Lee, Wonjong Lee +3Large Language Model-Based Robot Planning Robot Task Planning
Dept. of Electronic Engineering, Sogang University, Seoul, Korea
OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning
Oct 6, 2026Hyeongwoo Nam, Woongje Cho, Juwon Kim +1Large Language Model-Based Robot Planning Robot Task Planning
School of Mechanical Engineering, Yonsei University, Seoul, Republic of Korea

Feb 25, 2026J/K move · Enter open · S save

Hierarchical LLM-Based Multi-Agent Framework with Prompt Optimization for Multi-Robot Task Planning

Tomoya Kawabe, Rin Takano

Data Science Laboratories, NEC Corporation, 1753, Shimonumabe, Nakahara-ku, Kawasaki, Kanagawa, 211-8666, Japan.

Large Language Model-Based Robot Planning Multi-Agent LLM Systems Robotics Prompt Optimization Multi-Robot Motion Planning

Open paper AskPDF arXiv

Multi-robot task planning requires decomposing natural-language instructions into executable actions for heterogeneous robot teams. Conventional Planning Domain Definition Language (PDDL) planners provide rigorous guarantees but struggle to handle ambiguous or long-horizon missions, while large language models (LLMs) can interpret instructions and propose plans but may hallucinate or produce infeasible actions. We present a hierarchical multi-agent LLM-based planner with prompt optimization: an upper layer decomposes tasks and assigns them to lower-layer agents, which generate PDDL problems solved by a classical planner. When plans fail, the system applies TextGrad-inspired textual-gradient updates to optimize each agent's prompt and thereby improve planning accuracy. In addition, meta-prompts are learned and shared across agents within the same layer, enabling efficient prompt optimization in multi-agent settings. On the MAT-THOR benchmark, our planner achieves success rates of 0.95 on compound tasks, 0.84 on complex tasks, and 0.60 on vague tasks, improving over the previous state-of-the-art LaMMA-P by 2, 7, and 15 percentage points respectively. An ablation study shows that the hierarchical structure, prompt optimization, and meta-prompt sharing contribute roughly +59, +37, and +4 percentage points to the overall success rate.

Abstract

Explore similar work

Hierarchical LLM-Based Multi-Agent Framework with Prompt Optimization for Multi-Robot Task Planning

HEART: Coordination of Heterogeneous Expert Agents for Physically Grounded Robotic Task Planning

OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning