CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy
# CHORUS: Single AI Policy Enables Decentralized Multi-Robot Teams
Researchers have developed CHORUS, a system allowing multiple robots to coordinate tasks using a single shared AI policy rather than separate control systems for each unit. Each robot operates independently using the same foundation model—a Vision Language Action (VLA) policy—while receiving only local observations from its own sensors. The system was tested on collaborative tasks requiring multiple robots to work together, demonstrating that this approach scales across different team sizes without retraining the core policy.
The challenge this addresses is fundamental to multi-robot operations: traditional centralized systems that fuse all sensor data into one control center become computationally expensive and fragile as teams grow. Decentralized alternatives typically require separate policies per robot or extensive retraining when team composition changes. For logistics, construction, and warehouse automation, this matters because operational flexibility—swapping robots in or out, adjusting team sizes, or deploying to new environments—currently demands significant engineering overhead. CHORUS potentially reduces that barrier.
The practical implication worth noting: the system still relies on real-time communication between robots to coordinate actions, meaning network reliability and latency become operational constraints rather than just performance factors. Whether this dependency proves manageable in environments with spotty connectivity or high interference remains an open question that operators will need to evaluate for their specific settings.