Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MRBTP: Efficient Multi-Robot Behavior Tree Planning and Collaboration
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 MRBTP, a sound and complete multi-robot behavior tree planning algorithm with cross-tree expansion and intention sharing, plus an optional LLM-based subtree pre-planning plugin that speeds planning and execution.
Problem
Multi-robot behavior tree planning is hard because heterogeneous actions must be coordinated across different BTs and homogeneous actions must provide fault tolerance without redundant execution; planning can also be slow over fine-grained action spaces.
Approach
Cross-tree expansion lets a condition expanded in one robot's BT be expanded by all robots' BTs, backup structures are kept for homogeneous actions, and robots broadcast intentions so others predict effects and avoid redundant execution; an LLM optionally pre-plans long-horizon subtrees for each robot.
Key results
- MRBTP achieves 100% success rate across homogeneity settings in warehouse management, while BT-Expansion drops to 12.4% at alpha≈0.5 and 4.6% at alpha=0 (4 robots).
- Subtree pre-planning with feedback reduces planning time from timeout (no subtree) to 3.7 s under full heterogeneity (alpha=1) and to 1.4 s under alpha=0.
- Under alpha=1 with subtree and subtree intention sharing, team steps drop from 114.4 (no intention sharing) to 77.1 with feedback, showing the benefit of intention sharing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MRBTP with cross-tree expansion instead of independent BT-Expansion | 4-robot warehouse planning · Warehouse management, α ≈ 0.5 and α = 0 | SR (%): 12.4 -> 100; 4.6 -> 100 | Helped |
| Add MRBTP intention sharing | MRBTP · Warehouse management, α = 1 | TS: 8.8 -> 5.8 | Helped |
| Add MRBTP cross-tree expansion | Multi-robot BT planning · Warehouse management, α ≈ 0.5 and α = 0 | SR (%): 12.4 -> 100; 4.6 -> 100 | Helped |
| Add atomic-action intention sharing (ablation) | MRBTP · Home service, α = 1 | TS: 161 -> 159.4 | Helped |
| Add subtree pre-planning (ablation) | MRBTP · Home service, α = 1 | TS: 161 -> 114.4 | Helped |
| Use subtree intention sharing rather than atomic-action intention sharing (ablation) | MRBTP with subtree pre-planning · Home service, α = 1 | TS: 109.6 -> 78.9 | Helped |
| Add subtree intention sharing to subtree pre-planning (ablation) | MRBTP · Home service, α = 1 | TS: 114.4 -> 78.9 | Helped |
| Add feedback to subtree generation (ablation) | MRBTP with subtree pre-planning · Home service, α = 1 | PT (s): 12.4 -> 3.7 | Helped |
| Add feedback to subtree generation (ablation) | MRBTP with subtree pre-planning · Home service, α = 1 | TS: 78.9 -> 77.1 | Helped |
| Add feedback to subtree generation (ablation) | MRBTP with subtree pre-planning · Home service, α = 1 | RS: 217.4 -> 205.2 | Helped |
| Add feedback to subtree generation (ablation) | MRBTP with subtree pre-planning · Home service, α = 1 | EC: 998.1 -> 384.3 | Helped |
| Use GPT-4o instead of GPT-3.5-turbo for subtree generation (ablation) | MRBTP with subtree and subtree IS · Home service, α = 1, feedback | TS: 80.0 -> 74.9 | Helped |
Surprising findings
- Under full homogeneity, MRBTP still outperforms BT-Expansion in execution efficiency thanks to intention sharing.
- Communication overhead remains largely unchanged across GPT-3.5-turbo, GPT-4o-mini and GPT-4o, while execution efficiency improves slightly with stronger LLMs.
Limitations
- Assumes serial, priority-based execution order for soundness/completeness proofs; parallel execution may cause deadlocks or departures from the region of attraction.
- Does not consider action costs, so solutions are not optimal.
- Subtree generation relies on LLM availability; evaluated only in simulation.
Experimental setup
Datasets: Minigrid-based warehouse management environment VirtualHome home service environment
Benchmarks: Minigrid warehouse management VirtualHome home service
Metrics: success rate team steps total robot steps communication overhead expanded conditions planning time
Baselines: BT-Expansion
Builds on: BT-Expansion Minigrid VirtualHome
Closest papers by meaning
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.86)
- RoboPARA: Dual-Arm Robot Planning with Parallel Allocation and Recomposition Across Tasks (ICLR 2026 · similarity 0.85)
- Tree-Planner: Efficient Close-loop Task Planning with Large Language Models (ICLR 2024 · similarity 0.85)
- Tru-POMDP: Task Planning Under Uncertainty via Tree of Hypotheses and Open-Ended POMDPs (NeurIPS 2025 · similarity 0.84)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.84)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.83)
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.