A new paper introduces DuoMind, a distributed hierarchical framework that coordinates multiple robots using semantic communication. Each robot pairs a VLA-based action model with a VLM-based orchestrator that reasons over tasks, local observations, and peer messages. The authors also release RoboPoly, a benchmark of long-horizon multi-robot manipulation tasks, and report performance gains on RoboPoly and RoboTwin.
A new arXiv paper introduces COMPASS, a decentralized multi-robot architecture in which each robot generates local feedback via a spatial transformer aggregating multi-hop messages across the fleet into a learned feedback token. The authors report it scales to 1024 robots, generalizes zero-shot to ambiguous instructions, and outperforms a centralized frontier LLM policy and a language-only communication ablation. It targets the failure of LLM-based approaches on multi-robot tasks as team sizes grow.
A new arXiv paper introduces COMPASS, a decentralized multi-robot architecture in which each robot generates local feedback via a spatial transformer aggregating multi-hop messages across the fleet into a learned feedback token. The authors report it scales to 1024 robots, generalizes zero-shot to ambiguous instructions, and outperforms a centralized frontier LLM policy and a language-only communication ablation. It targets the failure of LLM-based approaches on multi-robot tasks as team sizes grow.
Researchers introduce AeroWeaver, a harness that integrates LLM agents into UAV swarms, weaving individual drone skills into coordinated mission-level behavior. It grounds semantic decisions in governed skills, runs role-conditioned local agents for distributed coordination, and refines skill selection online from accumulated experience. The system targets search, inspection, and tracking missions without a central agent.
Researchers introduce AeroWeaver, a harness that integrates LLM agents into UAV swarms, weaving individual drone skills into coordinated mission-level behavior. It grounds semantic decisions in governed skills, runs role-conditioned local agents for distributed coordination, and refines skill selection online from accumulated experience. The system targets search, inspection, and tracking missions without a central agent.
A new arXiv paper introduces a communication-efficient digital-twin coordination framework for teams of heterogeneous LLM-powered embodied agents operating under limited network resources. The approach addresses the overhead of multi-round natural-language inter-agent dialogue by leveraging digital twins to reduce communication rounds. The work targets applications in smart factories, warehouses, and service robotics.
A new arXiv paper introduces a communication-efficient digital-twin coordination framework for teams of heterogeneous LLM-powered embodied agents operating under limited network resources. The approach addresses the overhead of multi-round natural-language inter-agent dialogue by leveraging digital twins to reduce communication rounds. The work targets applications in smart factories, warehouses, and service robotics.
A new arXiv paper introduces CoAdapt, a framework using an LLM as a runtime fusion controller for collaborative perception in Industrial IoT robotic swarms. The LLM decides which robots join fusion and which algorithm to apply, no task-specific training needed. On the OPV2V benchmark it cuts communication cost by 38% while matching static baselines' detection precision.
A new arXiv paper introduces CoAdapt, a framework using an LLM as a runtime fusion controller for collaborative perception in Industrial IoT robotic swarms. The LLM decides which robots join fusion and which algorithm to apply, no task-specific training needed. On the OPV2V benchmark it cuts communication cost by 38% while matching static baselines' detection precision.
An arXiv paper presents a decentralized multi-robot navigation system where LLM policy reasoning operates at round level while local control runs at tick level. Each robot independently pairs an LLM policy agent with a UCB bandit and a Double DQN controller, with no central LLM coordinating team actions. The approach refines robot policies without destabilizing local learning, demonstrated in a NetLogo-Python implementation.
An arXiv paper presents a decentralized multi-robot navigation system where LLM policy reasoning operates at round level while local control runs at tick level. Each robot independently pairs an LLM policy agent with a UCB bandit and a Double DQN controller, with no central LLM coordinating team actions. The approach refines robot policies without destabilizing local learning, demonstrated in a NetLogo-Python implementation.