Research map / Multi-agent systems and games / LLM agents
CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation
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 CaPo, a two-phase LLM multi-agent framework that first generates and discusses a long-term meta-plan and then adapts it to agent progress during execution, improving transport rate and efficiency on TDW-MAT and C-WAH.
Problem
LLM-based embodied agents cooperating on joint tasks often act extemporaneously with short-term, single-step plans, causing redundant actions and failures in tasks like object transport and household activities.
Approach
Agents collaboratively create a long-term meta-plan through multi-turn discussion (designer proposes, teammates evaluate), then dynamically revise it via progress-triggered discussions when new objects are found or subtasks completed, instead of CoELA's per-step short-term planning.
Key results
- On TDW-MAT with GPT-4 agents and oracle perception, CaPo reaches 89% average transport rate versus CoELA's 85% and RoCo's 85%.
- On TDW-MAT with GPT-3.5 agents and oracle perception, CaPo achieves 84% transport rate versus CoELA's 72%.
- On C-WAH with GPT-4 agents (symbolic obs), CaPo needs 51 average steps versus CoELA's 57, a 9.8% relative improvement reported in the text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CaPo instead of CoELA | GPT-4 embodied agents · TDW-MAT w/ Oracle Perception, Avg. | Transport Rate (%): 85 -> 89 | Helped |
| Use CaPo instead of CoELA | GPT-3.5 embodied agents · TDW-MAT w/ Oracle Perception, Avg. | Transport Rate (%): 72 -> 84 | Helped |
| Use CaPo instead of CoELA | GPT-4 embodied agents · C-WAH Symbolic Obs | Average Steps: 57 -> 51 | Helped |
| Use CaPo instead of RoCo | GPT-4 embodied agents · C-WAH Visual Obs | Average Steps: 89 -> 83 | Helped |
| Add meta-plan initialization (ablation) | CaPo₁ on TDW-MAT · TDW-MAT Avg. | Transport Rate (%): 73 -> 74 | Helped |
| Add meta-plan evaluation and optimization (ablation) | CaPo₂ on TDW-MAT · TDW-MAT Avg. | Transport Rate (%): 74 -> 77 | Helped |
| Add progress-adaptive meta-plan (ablation) | CaPo₃ on TDW-MAT · TDW-MAT Avg. | Transport Rate (%): 77 -> 84 | Helped |
| Increase agent count from 2 to 3 (ablation) | CaPo on C-WAH · C-WAH Visual Obs | Average Steps: 83 -> 72 | Helped |
| Increase agent count from 3 to 4 (ablation) | CaPo on C-WAH · C-WAH Visual Obs | Average Steps: 72 -> 74 | Hurt |
| Increase discussion budget from 1 to 3 rounds (ablation) | CaPo on TDW-MAT · TDW-MAT | Transport rate (%): 79 -> 84 | Helped |
| Increase discussion budget from 3 to 4 rounds (ablation) | CaPo on TDW-MAT · TDW-MAT | Transport rate (%): 84 -> 84 | No clear effect |
| Increase discussion budget from 4 to 5 rounds (ablation) | CaPo on TDW-MAT · TDW-MAT | Transport rate (%): 84 -> 83 | Hurt |
Surprising findings
- Ablation shows the progress-adaptive meta-plan adds about 7 points over meta-plan generation alone with GPT-3.5 on TDW-MAT.
- Increasing agents from three to four on C-WAH yields only minor or degraded gains, attributed to frequent agent dispatch.
Limitations
- Relies heavily on LLM reasoning for meta-plan generation and adaptation.
- Uses a pre-defined communication protocol that may limit flexibility in dynamic or heterogeneous settings.
- Evaluation limited to two-agent simulated benchmarks with a small number of test episodes (24 TDW-MAT, 10 C-WAH).
Experimental setup
Datasets: TDW-MAT test set (24 episodes) C-WAH test set (10 episodes) RoCoBench tasks
Benchmarks: TDW-MAT C-WAH RoCoBench
Metrics: transport rate average steps efficiency improvement success rate consensus/failed-case rate human-rated communication effectiveness, helpfulness, trust
Baselines: CoELA ProAgent RoCo MHP RHP
Builds on: CoELA Watch-And-Help DyLAN
Closest papers by meaning
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.92)
- Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 2024 · similarity 0.90)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.90)
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.90)
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.89)
- REVECA: Adaptive Planning and Trajectory-Based Validation in Cooperative Language Agents Using Information Relevance and Relative Proximity (AAAI 2025 · similarity 0.89)
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.