Research map / Multi-agent systems and games / LLM agents
EMOS: Embodiment-aware Heterogeneous Multi-robot Operating System with LLM Agents
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 EMOS, an LLM multi-agent framework that gives each robot a self-generated "robot resume" from its URDF (LLM summary plus forward-kinematics numerics) for embodiment-aware task planning, and introduces the Habitat-MAS benchmark with four heterogeneous multi-robot household tasks.
Problem
Heterogeneous multi-robot systems need task planning and assignment that respects each robot's physical embodiment, but LLM-based multi-agent systems typically rely on human-assigned roles that ignore hardware-dependent capabilities.
Approach
Replace human-authored role descriptions with self-prompted robot resumes built from URDF files (skeleton summary plus forward-kinematics numerical capability data), combined with a centralized leader-planner that assigns subtasks and per-robot reflection agents that check feasibility, followed by decentralized function-calling action execution.
Key results
- EMOS achieves 37.82% success rate on Habitat-MAS versus 23.56% without numerical capability descriptions, 15.63% without robot resume, and 15.23% without group discussion (ablations on 519 episodes).
- Removing numerical forward-kinematics descriptions drops multi-floor rearrangement success from 13.46% to 3.85%.
- Sub-goal success rate for EMOS is 81.26% overall, versus 65.27% for the w/o robot resume ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add numerical capability descriptions to robot resumes (ablation) | EMOS · Habitat-MAS | Success rate: 23.56% -> 37.82% | Helped |
| Add numerical capability descriptions to robot resumes (ablation) | EMOS · Single-floor and multi-floor rearrangement | Success rate: 28.35% -> 9.20%; 13.46% -> 3.85% | Helped |
| Add textual robot capability summaries from URDF (ablation) | EMOS · Habitat-MAS | Success rate: 15.63% -> 37.82% | Helped |
| Add textual robot capability summaries from URDF (ablation) | EMOS · Navigation | Success rate: 37.37% -> 14.14% | Helped |
| Add textual robot capability summaries from URDF (ablation) | EMOS · Perception | Success rate: 52.94% -> 28.32% | Helped |
| Add centralized group discussion and task assignment (ablation) | EMOS · Habitat-MAS | Success rate: 15.23% -> 37.82% | Helped |
| Add centralized group discussion and task assignment (ablation) | EMOS · Habitat-MAS | Token usage: 36377 -> 80783 | Hurt |
| Add centralized group discussion and task assignment (ablation) | EMOS · Robot_Detect_Object | Sub-goal success rate: 70.80% -> 93.81% | Helped |
Surprising findings
- Without group discussion, the agent uses far fewer tokens (36377 vs 80783) but success drops to 15.23%, while removing robot resume yields the most simulation steps.
- Removing the numerical description leaves single-floor rearrangement success (9.20%) about the same as removing the whole robot resume (12.99%).
Limitations
- Ablations run on a 519-episode subset due to API budget constraints.
- Perfect SLAM and ground-truth semantic mesh are assumed; low-level control uses contact-based grasping with physics disabled.
- LLM hallucinations in spatial reasoning (e.g., incorrect reach computations) cause misassignments.
Experimental setup
Datasets: Habitat-MAS (HSSD scenes) Matterport3D
Benchmarks: Habitat-MAS
Metrics: success rate sub-goal success rate token usage simulation steps
Baselines: w/o. Numerical capability description w/o. Robot resume (role description) w/o. Group discussion
Builds on: HMAS-2 Habitat Hydra CAMEL MetaGPT
Closest papers by meaning
- Long-Horizon Planning for Multi-Agent Robots in Partially Observable Environments (NeurIPS 2024 · similarity 0.85)
- MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.85)
- RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.84)
- Debate2Create: Robot Co-design via Multi-Agent LLM Debate (ICML 2026 · similarity 0.84)
- OWL: Optimized Workforce Learning for General Multi-Agent Assistance in Real-World Task Automation (NeurIPS 2025 · similarity 0.84)
- OWMM-Agent: Open World Mobile Manipulation With Multi-modal Agentic Data Synthesis (NeurIPS 2025 · similarity 0.84)
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.