Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Long-Horizon Planning for Multi-Agent Robots in Partially Observable Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LLaMAR, a plan-act-correct-verify cognitive architecture of four LM modules for long-horizon multi-agent robot planning under partial observability, plus the MAP-THOR benchmark in AI2-THOR.
Problem
Long-horizon task planning for multiple embodied robots from natural language instructions in partially observable, dynamic environments, without oracle or simulator feedback for subtask verification or action correction.
Approach
Four LM roles (Planner, Actor, Corrector, Verifier) run in a centralized loop each high-level step; the Corrector uses execution failure feedback and images to suggest corrective actions, and the Verifier self-checks subtask completion, avoiding oracle information. A CLIP-guided heuristic handles exploration.
Key results
- LLaMAR with GPT-4V reaches 0.66 success rate on 2-agent MAP-THOR versus 0.34 for the best baseline ReAct (0.25 for CoELA, 0.11 for SmartLLM).
- Adding the Corrector with oracle feedback raises success from 0.45 to 0.67 in ablation; the full oracle-free system reaches 0.66.
- Success rate peaks at 3 agents (0.70 on MAP-THOR) and declines at 5 agents, while balance drops from 0.82 to 0.54.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LLaMAR instead of the strongest baseline ReAct | 2-agent MAP-THOR · MAP-THOR | success rate: 0.34 -> 0.66 | Helped |
| Add Planner and Verifier modules to the Actor (ablation) | LLaMAR · MAP-THOR, 2 agents | success rate: 0.33 -> 0.45 | Helped |
| Add Corrector with privileged oracle information (ablation) | LLaMAR · MAP-THOR, 2 agents | success rate: 0.45 -> 0.67 | Helped |
| Use the full method without privileged information rather than oracle-enhanced Corrector (ablation) | LLaMAR · MAP-THOR, 2 agents | success rate: 0.67 -> 0.66 | No clear effect |
| Increase the team from 2 to 3 agents | LLaMAR · MAP-THOR | success rate: 0.62 -> 0.70 | Helped |
| Increase the team from 3 to 5 agents | LLaMAR · MAP-THOR | success rate: 0.70 -> 0.62 | Hurt |
| Increase the team from 2 to 5 agents | LLaMAR · SAR | success rate: 0.44 -> 0.74 | Helped |
Surprising findings
- GPT-4 text-only performed worst among the tested LMs because it cannot reason about visual observations needed by the Corrector.
- Chain-of-thought performance declined with longer planning horizons, with success 0.14 versus 0.33 for plain Act.
Limitations
- Four LM queries per high-level step make it more expensive and slower than baselines.
- Limited spatial reasoning for obstacle avoidance and path planning.
- Performance bounded by the underlying VLM's instruction following and reasoning, with occasional misunderstanding of environment rules.
Experimental setup
Datasets: MAP-THOR
Benchmarks: MAP-THOR Search & Rescue (SAR)
Metrics: success rate transport rate coverage balance average steps
Baselines: Act Chain-of-Thought ReAct SmartLLM CoELA
Builds on: SmartLLM CoELA ReAct Chain-of-Thought CMAS
Closest papers by meaning
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.90)
- Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.89)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.89)
- Agent Planning with World Knowledge Model (NeurIPS 2024 · similarity 0.89)
- Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.88)
- Graph-enhanced Large Language Models in Asynchronous Plan Reasoning (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.